CrispASR

Environment variables

CrispASR exposes a large number of environment variables — tuning knobs, device placement overrides, benchmark timers, and debug/dump switches. This page is the single reference for all of them: the naming convention, the cross-cutting variables, the per-backend variables, and the ones CrispASR deliberately does not own.

This document was introduced for issue #265 (“consistency and documentation”). The same change standardized every backend-owned variable onto one prefix and unified the reference-voice cache location (see below).

Naming convention

Every CrispASR-owned variable is named:

CRISPASR_<BACKEND>_<FEATURE>

for example CRISPASR_OMNIVOICE_CODEC_GPU, CRISPASR_IRODORI_CODEC_GPU, CRISPASR_PARAKEET_BENCH. Process-global variables that are not tied to one backend drop the <BACKEND> segment (CRISPASR_MODELS_DIR, CRISPASR_KV_ON_CPU).

Legacy aliases and deprecation warnings

Historically many backends used a bare prefix (OMNIVOICE_CODEC_GPU, QWEN3_TTS_BENCH, CHATTERBOX_DEBUG, …). Those bare names still work as legacy aliases so existing scripts, notebooks, and Kaggle A/B kernels keep running. The alias is derived automatically by stripping the CRISPASR_ prefix, so for any CRISPASR_FOO_BAR documented here the bare FOO_BAR is also honored.

The first time a run reads a value from a legacy name, CrispASR prints a one-time notice to stderr:

[crispasr] warning: environment variable 'OMNIVOICE_CODEC_GPU' is deprecated and
will be removed in a future release; use 'CRISPASR_OMNIVOICE_CODEC_GPU' instead
(the old name still works for now). ...

Migrate to the canonical CRISPASR_-prefixed name at your convenience. To silence the warnings (e.g. for a pipeline that scrapes stderr and cannot migrate yet), set:

CRISPASR_SUPPRESS_ENV_DEPRECATION=1

The lookup + aliasing is implemented once in src/core/crispasr_env.h (crispasr_env::get / truthy / present); new backends should read env vars through that helper with a canonical CRISPASR_-prefixed name.

Value/truthiness conventions

Most variables are read with one of two conventions — check the surrounding suffix legend, but as a rule:

• Flag (opt-in) — set to 1 (or any non-empty, non-0 value) to enable. Some legacy flags treat being set at all as on, so prefer =1 / unset it entirely rather than =0 when a flag’s exact semantics matter.
• Path / value — the variable’s value is a filesystem path, integer, or float used directly.

Common suffix legend

The same feature suffixes recur across nearly every backend. Rather than repeat them per backend below, they mean:

Suffix	Meaning
_BENCH	Print per-stage wall-clock timings for that engine.
_DEBUG / _DIAG	Verbose debug logging.
_DUMP_DIR / _DUMP / _DUMP_*	Write intermediate tensors / stages to disk (diagnostics, diff-harness).
_FORCE_SCALAR	Disable SIMD kernels (numeric-parity debugging).
_GPU / _USE_GPU / _FORCE_METAL	Force that stage/codec onto the GPU.
_CPU / _CPU_ONLY / _FORCE_CPU	Force that stage/codec onto the CPU.
_NO_FA / _FLASH_ATTN	Toggle flash-attention.
_FASTCONV (_DEBUG)	Enable the baked-F32 / matmul conv fast path for a codec/vocoder.
_CFG_INTERVAL (_DEBUG)	Interval-CFG cadence for a flow-matching decoder.
_SEED	RNG seed for reproducible generation.
_*_FILE / _*_PATH / _*_GGUF	Override an asset/model/fixture path (mostly dev/test).

Global / cross-cutting variables

These are not tied to a single backend.

Paths, cache, and models

Variable	Purpose
CRISPASR_MODELS_DIR	Directory searched for GGUF models (also the auto-download target root).
CRISPASR_CACHE_DIR	Base cache directory for auto-downloaded models/assets. Resolution order is CRISPASR_CACHE_DIR → CRISPASR_MODELS_DIR → platform default ($HOME/.cache/crispasr; %USERPROFILE%\.cache\crispasr on Windows).
CRISPASR_SCRATCH_DIR	Scratch directory for temporary run artifacts.
CRISPASR_DUMP_DIR	Global tensor-dump directory (diagnostics).
CRISPASR_GGUF_MMAP / CRISPASR_GGUF_PRELOAD	Control GGUF mmap vs. preload-into-RAM loading.
CRISPASR_GGUF_MAX_ALLOC_CHUNK	Lower the per-buffer allocation cap in bytes (default 1.5 GiB) used to split a model across several backend buffers on drivers with a small maxMemoryAllocationSize (#276).
CRISPASR_MLOCK	mlock model weights into RAM.
CRISPASR_IGNORE_CPU_ISA	=1 exactly (no other value works) continues past the startup build-vs-host CPU instruction-set check (#380) instead of exiting; the process will SIGILL at the first compute op the CPU can’t run.

GPU / device placement

Variable	Purpose
CRISPASR_N_GPU_LAYERS	Number of transformer layers to offload to the GPU.
CRISPASR_ARG_DEVICE	Default device selection for the CLI.
CRISPASR_KV_ON_CPU	Keep the KV cache on the CPU.
CRISPASR_KV_QUANT / _KV_QUANT_K / _KV_QUANT_V / _KV_READ_F32	KV-cache quantization / read format.
CRISPASR_GPU_PREF_CPU_LEGACY	1 restores the pre-T18 behaviour where --gpu-backend cpu fell through to the best GPU. Off by default: the fall-through silently constructed a Metal device the flag exists to avoid.
CRISPASR_METAL_PIPELINE_CACHE_MAX_MB	Size cap in MiB (default 64) above which the on-disk Metal pipeline-cache archive is skipped at init — opening it costs ~1 ms/MB. 0 = uncapped (always use it). macOS only.

Device selection across compiled backends also honors the standard ggml / CUDA variables CUDA_VISIBLE_DEVICES and GGML_VK_VISIBLE_DEVICES — see “Variables CrispASR does not own” below.

Session / long-audio chunking

Variable	Purpose
CRISPASR_SESSION_AUTOCHUNK	Enable auto-chunking in the session API for long audio.
CRISPASR_SESSION_CHUNK_SECONDS	Chunk length (seconds) for session auto-chunking.
CRISPASR_SESSION_PERBACKEND_CHUNK	Use per-backend chunk-window tuning instead of a flat window.
CRISPASR_SESSION_UNIFIED_DISPATCH	Route surfaces through the unified library dispatch path.
CRISPASR_SLICE_PIPELINE	Force the CLI’s encode ∥ decode slice pipeline on/off. The override may only turn it OFF, or ON where it is already safe — it can never switch off one of the safety conditions (-p N, --return-logits, gap-fill re-entry, single slice).

Post-decode hygiene (PLAN.md §W2–W7)

All OFF unless set. Each of these can delete or alter user-visible text, so none of them switches on by surprise; a wrong deletion is worse than a surviving artifact. Applied on both the CLI and the session C-ABI.

Variable	Purpose
CRISPASR_SEG_MAX_CHARS	Truncate any segment longer than N code points, backing up to the last 。．.！!？?、, but never below 75% of the cap. A line past the cap is almost always a repetition hallucination that survived the n-gram collapse.
CRISPASR_SEG_DROP_NONVERBAL	Drop segments that are entirely a non-verbal marker — [Music], (applause), （喘ぎ声）, ♪. Running speech merely containing such a word is never dropped.
CRISPASR_SEG_LOGPROB_THOLD	Drop segments whose average log-probability is below this. Post-hoc, on top of the decoder’s own fallback gate.
CRISPASR_SEG_LOGPROB_MARGIN	Loosen that threshold by this much for segments ≤1.6 s — a short segment’s mean logprob is noisier, so it gets more room, not less.
CRISPASR_SEG_MERGE_REPEATS	Collapse runs of near-identical adjacent segments into one spanning the whole run. Catches a phrase repeating across segment boundaries, which per-segment loop fixes cannot see.
CRISPASR_SEG_MERGE_SIMILARITY	Similarity bar for the above (default 0.90; LCS over code points).
CRISPASR_SEG_MERGE_GAP_CS	Never merge across a gap wider than this many centiseconds (default 200). Two identical lines a minute apart are two real utterances.
CRISPASR_SEG_MERGE_MIN_RUN	Minimum consecutive similar segments before merging (default 3), so an ordinary repeated “yes.” pair survives.

Alignment and VAD sanity checks

Variable	Purpose
CRISPASR_ALIGN_SENTINEL	0 disables the forced-alignment collapse check. On by default, detect + warn only. Catches ctc_forced_align() returning words at t0 == t1 == 0 — its two silent-zero paths (characters absent from the CTC vocab, or a word the Viterbi path never visited) produce garbage timestamps inside a successful return.
CRISPASR_ALIGN_SENTINEL_REDISTRIBUTE	1 opts into repair: respace the words across the clip in proportion to character count. Off by default — a wrong auto-repair would be just as invisible as the collapse.
CRISPASR_VAD_FAILOVER	0 disables the VAD sanity check. On by default: if a clip over 120 s comes back with under 1% speech coverage (or a couple of segments covering under 10% of a very long clip), the VAD is wrong and the run falls back to fixed full-clip chunks rather than losing the transcript.
CRISPASR_NGRAM_LOOPFIX_OFF	1 disables the repeated-n-gram collapse entirely, exposing the RAW decoded text. Diagnostic: for telling whether a loop originates in the decode itself or is merely being masked.
CRISPASR_ORDER_WARN	0 disables the one-shot “segment timestamps go backwards” warning. On by default; detect + warn only. Cues that merely overlap are deliberately not flagged (gap-fill jitter).
CRISPASR_ALIGN_NO_ROMANIZE	1 passes non-Latin reference text through raw instead of auto-romanizing it for a CTC aligner with a Latin vocabulary (#252). Since #419 the romanization is only the aligner’s internal label — aligned words hand back the original script, so srt/vtt/-sp/-sow output no longer flips Cyrillic/CJK transcripts to transliteration.
CRISPASR_ALIGN_DEBUG	1 prints the romanized reference transcript the aligner actually used.

Decoding / beam search (shared)

Variable	Purpose
CRISPASR_MAES_BETA / _MAES_GAMMA / _MAES_NUM_STEPS	MAES beam-search parameters.
CRISPASR_TDT_BATCH / CRISPASR_RNNT_BATCH	Batch the TDT / RNNT joint decode.
CRISPASR_RNNT_GGML_PERSTEP	Per-step (vs. persistent-graph) ggml RNNT decode.
CRISPASR_RNNT_GPU_ENC_PROJ	Parakeet’s backend encoder-to-joint projection is default on CUDA. 0 restores the scalar CPU projection; 1 opts other GPU backends in.
CRISPASR_NGRAM_LOOPFIX_OFF	Disable the n-gram decode-loop breaker.
CRISPASR_STREAM_SLICE_MEMO	Memoize per-slice streaming partial decodes by absolute sample range (#404). Default ON — finals byte-equal, wall −12 % CPU / −6 % GPU in the quiet-box A/B; =0 re-decodes closed slices every step.
CRISPASR_GAP_FILL / _GAP_FILL_MIN_CS	Re-transcribe spans a first pass left empty (long audio); on by default for parakeet, threshold non-JA 300 cs / JA 100 cs.

G2P / phonemizer

Variable	Purpose
CRISPASR_CMUDICT_PATH	Path to the CMUdict pronunciation dictionary.
CRISPASR_DE_DICT_PATH / _FR_DICT_PATH / _ES_DICT_PATH / _RU_DICT_PATH	Language-specific pronunciation dictionaries.
CRISPASR_RU_HETERONYMS_PATH	Path to the Russian heteronym list. Only read when CRISPASR_G2P_RU_HETERONYM_WARN is on — it is diagnostic data, not a lookup tier.
CRISPASR_G2P_RU_HETERONYM_WARN	1 prints one line per input word that the upstream project flagged as genuinely ambiguous. Those 17,359 words are DISJOINT from the 812,953-entry vocabulary — they were removed because upstream could not choose a reading — so they have no dictionary entry and the letter-to-sound rules pick one reading from spelling alone. Common words are in that set (все, уже, потом, чем, небо, тест). Off by default; on, it turns a mispronunciation that looks like a rule bug into a named, explained limitation. Writing the ё or an explicit combining acute (замо́к) in the input resolves many of them, and both are honoured.
CRISPASR_G2P_RU_ANALOGY	0 disables the Russian stress-analogy tier, which finds a known relative of an OOV word (strip up to 3 letters, glue a short list of inflectional endings back on) and borrows its stress. On by default and measured on 10,000 held-out dictionary words: stressed-syllable index right 93.9% with it and 47.1% without; exact IPA match 79.4% vs 40.9%. The lever exists because it is the one tier that can take a stress from a word that merely LOOKS related.
CRISPASR_G2P_DICT_SOURCE / _G2P_MODEL_PATH	G2P dictionary source / neural G2P model path.
CRISPASR_ESPEAK_DATA_PATH	eSpeak-NG data directory.
CRISPASR_MISAKI_DICT_PATH	Path to the misaki US contextual-word dictionary (default ~/.cache/crispasr/misaki-us.txt) used by the English misaki G2P (#316).
CRISPASR_KOKORO_G2P	Kokoro G2P backend selection.
CRISPASR_KOKORO_MISAKI_IPA	0 disables the espeak-IPA → misaki-alphabet conversion Kokoro needs (#316), restoring the raw G2P spelling for A/B. On by default.
CRISPASR_G2P_DE_UNSTRESS	1 reads the German closed class the way espeak reads it in a SENTENCE (sie → ziː) instead of the citation form our per-word dictionary stores (zˈiː). Off by default: it takes phoneme agreement with espeak from 45.9% to 87.1%, but the ASR round-trip could not resolve a difference, and that metric measures intelligibility rather than naturalness (#316).
CRISPASR_KOKORO_DE_MISAKI_ALPHABET	1 applies misaki’s tied-sequence collapse for German (tsvˈaɪ → ʦvˈI), which is what the published training recipe does. Off by default: it made the ASR round-trip worse on the kokoro-de-hui-base model we ship, which appears to predate that part of the recipe (#316). The ʏ→y vocabulary fixup is applied either way.
CRISPASR_T5_REPEAT_BREAK	0 disables the decode-loop break for madlad/T5 translation, restoring exact PyTorch-blueprint behaviour. On by default: MADLAD greedy-decodes into a repeated token cycle on some short inputs and burns the whole token budget on it. The blueprint does the same — this is a deliberate improvement on it, not a parity fix (#333).
CRISPASR_T5_NO_KV_REUSE	1 re-forwards the whole decoder prefix each step instead of appending to the KV cache. Same output, much slower; an A/B lever for isolating cache bugs (#333).
CRISPASR_T5_KEEP_EMBED	1 keeps shared.embed.* and lm_head.* at source precision when quantizing a T5 model. Off by default because it was measured and loses: on madlad400 it makes q8_0 3.38→3.62 GB and q4_k 2.04→2.41 GB for a worst-stage cosine that does not improve (0.999922→0.999920, 0.992929→0.992606). The Q4_K loss accumulates through the 32 encoder blocks, not in the embedding lookup (#333).
CRISPASR_KOKORO_PUNCT	0 drops punctuation from the phoneme string for the German/French/Spanish built-in G2Ps, restoring pre-0.8.26 behaviour for A/B. On by default: Kokoro’s vocabulary contains ,.;:!? and they are how it pauses (#316). English is not gated — it is settled against misaki.

Watermark / provenance

Variable	Purpose
CRISPASR_NO_WATERMARK	Disable the audio watermark.
CRISPASR_WATERMARK_LEGACY	Use the legacy watermark path.
CRISPASR_WATERMARK_DETECT	Which statistic --detect-watermark uses: frames (per-frame t + decoy specificity) or sign / 0 (the older averaged-spectrum sign test, kept for A/B and for re-reading an older release’s score).
CRISPASR_NO_C2PA_REMUX	Skip the C2PA MP4 remux step.
CRISPASR_CONSENT_LOG	Path to a JSON-Lines sink for voice-cloning consent records. Without it the records only go to stderr, which is interleaved with model-load noise and so a poor evidential artefact.

Quantization / diff-harness / misc

Variable	Purpose
CRISPASR_QUANT_LMHEAD	Quantize the LM head during crispasr-quantize.
CRISPASR_IMATRIX_OUT	Importance-matrix output path.
CRISPASR_ACTDUMP_OUT / _ACTDUMP_TENSOR	Activation dump output / target tensor.
CRISPASR_DIFF_NO_GPU / _DIFF_USE_GPU / _DIFF_SLICES / _DIFF_STAGES	crispasr-diff harness controls.
CRISPASR_MEL_SERIAL	Force the serial STFT. The parallel mel/STFT path is DEFAULT ON since #305; this is the opt-out. (The older opt-in CRISPASR_MEL_PARALLEL is no longer read.)
CRISPASR_MEL_TIMING	Print mel/STFT stage timings.
CRISPASR_HQ_RESAMPLE	0 selects the cheap linear resampler for CLI input decoding instead of the high-quality one.
CRISPASR_CORE_ATTN_EAGER_F32	Force the shared attention helper’s eager (non-flash) path to F32 accumulation.
CRISPASR_CORE_ATTN_DUMP_FA_LAYER	Dump the shared attention helper’s flash-attn inputs/outputs for one layer index.
CRISPASR_VERBOSE	Global verbose output.
CRISPASR_NO_WARMUP / CRISPASR_WARMUP	Skip / force the model warmup pass.
CRISPASR_SERVER_WORKERS / CRISPASR_API_KEYS	HTTP server worker count / API keys.
CRISPASR_TEST_STREAM_THROW	Test-only: lets the server’s streaming worker throw on the magic input __throw_test__ (both the variable and the input are required, so it cannot fire in production).

Container / launcher

Read by the Docker images and .devops/run-server.sh, not by the C++ itself — crispasr-diagnostics echoes them so a support dump shows how the container was started.

Variable	Purpose
CRISPASR_BACKEND	Backend the container’s entrypoint should select (docker-compose*.yml, .env.example).
CRISPASR_USE_CUDA_COMPAT	1 prepends /usr/local/cuda/compat to LD_LIBRARY_PATH in the CUDA images — for hosts whose driver is older than the image’s CUDA runtime.

Vendored ggml (CrispASR-added)

CrispASR’s in-tree ggml/ carries a few CrispASR-prefixed knobs on top of upstream’s GGML_* set. They are read by getenv directly (no legacy alias).

Variable	Purpose
CRISPASR_GGML_ALLOC_TRACE	Trace ggml-alloc graph-allocation decisions.
CRISPASR_GGML_ALLOC_TRACE_MAX_PASSES	Number of allocation passes that trace prints (default: unlimited once the trace is on).
CRISPASR_METAL_N_CB	Override the Metal backend’s command-buffer count (#83).
CRISPASR_METAL_PROFILE	1 whole-graph host/GPU split, 2 per-op breakdown, 3 per-op plus a per-node trace announced before each encode — with 3 the last line names the node an encode faulted on.
CRISPASR_METAL_STRICT_FP	Compile the Metal kernels with fast-math OFF (#83). Costs throughput; buys bit-identical CPU/GPU output where operand reordering was downconverting F32 intermediates.
CRISPASR_METAL_FORCE_BARRIER	1 forces a memory barrier before every Metal op (concurrency-hazard bisection, #83).
CRISPASR_METAL_IM2COL_FLAT	0 restores the legacy IM2COL Metal kernel. The flat kernel (one thread per dst element) is the default.
CRISPASR_FORCE_BLIT_COPY	Use the blit-encoder copy path even for a shared (unified-memory) Metal buffer (#83).
CRISPASR_FORCE_DMB	Insert a full memory barrier after the host memcpy into a shared Metal buffer (#83).

Reference-voice cache (voice cloning)

Encoding a reference clip for voice cloning (a codec encoder, a Conformer / Perceiver, or an ASR pass) is slow and produces a small, reusable blob. Every voice-cloning TTS backend caches that blob through one shared mechanism (src/core/tts_ref_cache.h), so the location and disable switch are identical across backends (issue #265 unified OmniVoice — which previously used a bespoke ~/.cache/crispasr cache — onto this path):

Variable	Purpose
CRISPASR_TTS_REF_CACHE	Set to 0 to disable reference-voice caching everywhere.
CRISPASR_TTS_REF_CACHE_DIR	Override the cache directory (default <TMPDIR>/crispasr-tts-refcache).

Content-addressed entries are keyed by a hash of the raw reference (plus, for some backends, an encoder-weight fingerprint) and tagged per backend (irodori-latent, openvoice2-se, omnivoice-voice, f5-reftext, …) so no two backends read each other’s blob. OmniVoice additionally honors the legacy CRISPASR_OMNIVOICE_VOICE_CACHE=0 as an alias for the shared disable switch.

Variables CrispASR does not own

These are read by CrispASR but are OS / third-party conventions and are intentionally not renamed to the CRISPASR_ prefix:

Variable	Origin
HOME, USERPROFILE, LOCALAPPDATA, XDG_CACHE_HOME	OS home / cache dirs.
TMPDIR, TEMP, TMP	OS temp dir (also the default ref-cache root).
HF_TOKEN, HUGGING_FACE_HUB_TOKEN	HuggingFace auth (Hub convention).
CUDA_VISIBLE_DEVICES	CUDA device selection.
GGML_VK_VISIBLE_DEVICES	ggml Vulkan device selection.
LLAMA_*	Vendored llama.cpp (talk-llama example).

Sibling modular libraries (own conventions)

The in-tree modular libraries keep their own established prefixes and are synced from their sibling repos, so their variables are not part of the CRISPASR_ scheme:

• crisp_audio/ — audio tower: CRISP_AUDIO_DUMP_STAGES, CRISP_AUDIO_KEEP_PAD_FRAMES, CRISP_AUDIO_WINDOWED_ATTN.
• glint/ — clean-room MP3/AAC codec: GLINT_*, AACDBG.
• crisp_lid/ · crisp_punc/ · crisp_truecase/ — standalone LID / punctuation / truecasing libraries; they and their src/ counterparts follow each library’s own naming (LID_*, FIREREDPUNC_*, PCS_*, TRUECASER_*). FIREREDPUNC_DEBUG=1 prints each restore pass as [PUNCDBG] in=<…> / out=<…>, which is the quickest way to see a backend’s true model output — --no-punctuation is not, because it strips punctuation after the fact and so hides text the model punctuated itself.

  ⚠ These libraries are built by the main target, and each has a second copy under src/. src/CMakeLists.txt prefers crisp_punc/ (etc.) and falls back to the src/ copy only when the sibling directory is missing from a checkout — so the crisp_punc/ copy is what normally links, and a change must be applied to both. #308’s capitalisation fix went into src/fireredpunc.cpp alone and was dead code for months while the shipping copy kept the bug. tests/test-punc-copies-in-sync.cpp now fails if they diverge.

Test fixtures

The live/integration test suite reads a few CRISPASR_-prefixed fixture variables (model/audio paths); tests/env-live-tests.sh sets sensible defaults. Like every other variable here, the pre-standardization bare names are still honored as deprecated aliases:

Variable	Purpose
CRISPASR_PARAFORMER_MODEL / _PARAFORMER_MODEL_Q4K / _PARAFORMER_AUDIO_ZH / _PARAFORMER_AUDIO_EN	Paraformer test model / audio paths.
CRISPASR_AUDIOSEAL_GGUF	AudioSeal test model path.
CRISPASR_PIPER_TEST_MODEL	Piper phonemize test model path.
CRISPASR_OV2_DUMP_DIR	OpenVoice2 HiFi test dump directory.
CRISPASR_MODEL_BTC_CHORDS	BTC chord-recognition test model path. Defaults to $CRISPASR_MODELS_DIR/btc-chords-large-f32.gguf.

---

Per-backend variables

Every variable below is the canonical CRISPASR_-prefixed name; the bare form (without CRISPASR_) is accepted as a deprecated alias. See the suffix legend above for the meaning of the common _BENCH / _DEBUG / _DUMP_* / _GPU / _CPU / _FORCE_SCALAR / _FASTCONV / _CFG_INTERVAL / _SEED suffixes.

AAC codec

• CRISPASR_AAC_DEBUG
• CRISPASR_AAC_DECODER

ARK-ASR

• CRISPASR_ARKASR_BLOCK_FROM_ID
• CRISPASR_ARKASR_CPU
• CRISPASR_ARKASR_DEBUG_GEN
• CRISPASR_ARKASR_GPU
• CRISPASR_ARKASR_INSTRUCTION
• CRISPASR_ARKASR_MAX_SINGLE_PASS_S
• CRISPASR_ARKASR_NO_CHUNK_CONTEXT
• CRISPASR_ARKASR_NO_EOS_SUPPRESS
• CRISPASR_ARKASR_NO_SPECIAL_SUPPRESS
• CRISPASR_ARKASR_TIMING

AudioSeal watermark

• CRISPASR_AUDIOSEAL_BENCH
• CRISPASR_AUDIOSEAL_DEBUG
• CRISPASR_AUDIOSEAL_DUMP_STAGES

BananaMind TTS

• CRISPASR_BANANAMIND_DEBUG
• CRISPASR_BANANAMIND_TTS_BENCH

Bark

• CRISPASR_BARK_BENCH
• CRISPASR_BARK_DECODE_CODES
• CRISPASR_BARK_DUMP_DIR

Beat-This (beat tracking)

• CRISPASR_BEAT_THIS_DEBUG

BERT encoder

• CRISPASR_BERT_ENCODER_BENCH

BTC chord recognition

• CRISPASR_BTC_DEBUG
• CRISPASR_BTC_DUMP_FEAT
• CRISPASR_BTC_MAJ_MIN — collapse the 170-class chord output to the 25-class maj/min vocabulary. Default off (full 170-class output): 170 reduces to maj/min at runtime, but a 25-class model can never be expanded, so the 170-class checkpoint is the shipping default.

The shipped BTC weights are CC-BY-NC-SA (trained on Isophonics / Robbie Williams / UsPop2002 chord annotations) even though the upstream BTC code and CrispASR itself are MIT. The registry refuses to download them without --accept-license cc-by-nc-sa-4.0 (or CRISPASR_ACCEPT_LICENSE). A commercial product must supply its own weights.

Canary / Canary-CTC / Canary-Qwen

• CRISPASR_CANARY_BENCH
• CRISPASR_CANARY_CTC_BENCH
• CRISPASR_CANARY_QWEN_BENCH
• CRISPASR_CANARY_QWEN_DEBUG
• CRISPASR_CANARY_QWEN_MIN_ENC_FRAMES
• CRISPASR_CANARY_QWEN_NO_ECHO_STRIP
• CRISPASR_CANARY_LEGACY_STREAM
• CRISPASR_CANARY_SEAM_DEDUP
• CRISPASR_CANARY_STREAM_THRESHOLD_S

Chatterbox

• CRISPASR_CHATTERBOX_BENCH
• CRISPASR_CHATTERBOX_BENCH_B2
• CRISPASR_CHATTERBOX_DEBUG
• CRISPASR_CHATTERBOX_DUMP_ATTN_AT
• CRISPASR_CHATTERBOX_DUMP_FFN_AT
• CRISPASR_CHATTERBOX_DUMP_GPT2_LAYERS
• CRISPASR_CHATTERBOX_DUMP_KPROJ_AT
• CRISPASR_CHATTERBOX_DUMP_KROPE_AT
• CRISPASR_CHATTERBOX_DUMP_KV_AT
• CRISPASR_CHATTERBOX_DUMP_KV_LAYER
• CRISPASR_CHATTERBOX_DUMP_LAYER
• CRISPASR_CHATTERBOX_DUMP_LOGITS_AT
• CRISPASR_CHATTERBOX_DUMP_NORM_AT
• CRISPASR_CHATTERBOX_DUMP_QPROJ_AT
• CRISPASR_CHATTERBOX_DUMP_VPROJ_AT
• CRISPASR_CHATTERBOX_DUMP_WK
• CRISPASR_CHATTERBOX_FLASH_ATTN — force ggml_flash_attn_ext for the T3 GPT-2 (turbo/nano) attention even on Vulkan, where naive attention is the default since issue #402 (RADV 780M crashes in the Vulkan FLASH_ATTN_EXT pipeline; the explicit softmax(QK^T)V path is verified working there).
• CRISPASR_CHATTERBOX_FORCE_GPU
• CRISPASR_CHATTERBOX_FULL_CPU
• CRISPASR_CHATTERBOX_LANG
• CRISPASR_CHATTERBOX_KV_CONT — materialize GPT-2 K/V layer views before flash attention, reproducing the pre-PR-410 path for correctness/performance A/Bs. The default passes the views directly; naive attention still materializes them because its matrix operations require that layout.
• CRISPASR_CHATTERBOX_NAIVE_ATTN — force the explicit softmax(QK^T)V T3 attention on every backend (debug gate; outranks _FLASH_ATTN).
• CRISPASR_CHATTERBOX_S3GEN_CPU
• CRISPASR_CHATTERBOX_T3_CPU_S3GEN_GPU
• CRISPASR_CHATTERBOX_T3_GPU
• CRISPASR_CHATTERBOX_SEED
• CRISPASR_CHATTERBOX_SYN_TEXT
• CRISPASR_CHATTERBOX_T3_BUCKET_REUSE
• CRISPASR_CHATTERBOX_T3_CFG_B2
• CRISPASR_CHATTERBOX_T3_CFG_BUCKET
• CRISPASR_CHATTERBOX_T3_SEED
• CRISPASR_CHATTERBOX_TEMP
• CRISPASR_CHATTERBOX_THREADS

Chatterbox S3Gen

• CRISPASR_S3GEN_CFG_INTERVAL
• CRISPASR_S3GEN_DUMP
• CRISPASR_S3GEN_DUMP_UNET
• CRISPASR_S3GEN_DUMP_UNET_NO_AUTO_MARK
• CRISPASR_S3GEN_ENCODER_CPU
• CRISPASR_S3GEN_FASTCONV
• CRISPASR_S3GEN_FASTCONV_DEBUG
• `CRISPASR_S3GEN_SIMDCONV` — opt into the CPU HiFT packed SIMD path for all 72 ResBlock Conv1d kernels; default off. Stays OPT-IN deliberately: the Kaggle A/B measured a 5.4 % s3gen REGRESSION on Xeon avx512f against the contributor’s 1.25x win on Zen 4 — micro-arch dependent; A/B on your own hardware before enabling.
• CRISPASR_S3GEN_SIMDCONV_DEBUG — print pack count, selected ISA, and CPU/GPU fallback status.
• CRISPASR_S3GEN_RC_AS_MUL_MAT
• CRISPASR_S3GEN_VOCODER_CPU
• CRISPASR_S3GEN_UNET_CFG_SINGLE
• CRISPASR_S3GEN_UNET_CPU
• CRISPASR_S3GEN_UNET_GALLOCR
• CRISPASR_S3GEN_UNET_KEEP_GPU_OP
• CRISPASR_S3GEN_UNET_MARK_DB_OUT
• CRISPASR_S3GEN_UNET_MARK_DB_RESNET
• CRISPASR_S3GEN_UNET_MARK_DB_TB
• CRISPASR_S3GEN_UNET_MARK_MB_OUT
• CRISPASR_S3GEN_UNET_MARK_MB_OUT_INDEX
• CRISPASR_S3GEN_UNET_MARK_MB_OUT_MAX
• CRISPASR_S3GEN_UNET_MARK_MB_RESNET
• CRISPASR_S3GEN_UNET_PIN_CPU_OP
• CRISPASR_S3GEN_UNET_PRESERVE_INTERMEDIATES
• CRISPASR_S3GEN_UNET_PROBE_BLOCK1
• CRISPASR_S3GEN_UNET_PROBE_DENOISER_OUT
• CRISPASR_S3GEN_UNET_PROBE_INPUT_SNAPSHOT
• CRISPASR_S3GEN_UNET_PROBE_RC_OUT

Chatterbox sub-modules

• CRISPASR_CB_CAMPPLUS_BENCH
• CRISPASR_CB_S3GEN_BENCH
• CRISPASR_CB_S3TOK_BENCH
• CRISPASR_CB_VE_BENCH

Cohere

• CRISPASR_COHERE_BENCH
• CRISPASR_COHERE_DEBUG
• CRISPASR_COHERE_DEVICE
• CRISPASR_COHERE_DUMP_ATTN
• CRISPASR_COHERE_DUMP_ENCOUT
• CRISPASR_COHERE_DUMP_MEL
• CRISPASR_COHERE_DUMP_STAGES
• CRISPASR_COHERE_FLASH
• CRISPASR_COHERE_GAPS
• CRISPASR_COHERE_LANGS
• CRISPASR_COHERE_LEGACY_SA
• CRISPASR_COHERE_PROBE_MAX_LANGS
• CRISPASR_COHERE_PROBE_REUSE_ENC
• CRISPASR_COHERE_PROBE_TEXTLID
• CRISPASR_COHERE_PROF
• CRISPASR_COHERE_SILENCE_GATE
• CRISPASR_COHERE_THREADS

Confucius4 TTS

• CRISPASR_CONFUCIUS4_BEAMS
• CRISPASR_CONFUCIUS4_CFG_FUSE
• CRISPASR_CONFUCIUS4_CFG_RATE
• CRISPASR_CONFUCIUS4_COND_DIR
• CRISPASR_CONFUCIUS4_COND_PYEMB
• CRISPASR_CONFUCIUS4_DUMP_S2A
• CRISPASR_CONFUCIUS4_GRAPH_EMBED
• CRISPASR_CONFUCIUS4_LR_LEGACY
• CRISPASR_CONFUCIUS4_MAX_LAYERS
• CRISPASR_CONFUCIUS4_PERSIST
• CRISPASR_CONFUCIUS4_REP_PEN
• CRISPASR_CONFUCIUS4_S2A_TEMP
• CRISPASR_CONFUCIUS4_SCHED
• CRISPASR_CONFUCIUS4_TEXT_IDS
• CRISPASR_CONFUCIUS4_T_SCHEDULE

CosyVoice3

• CRISPASR_COSYVOICE3_BENCH
• CRISPASR_COSYVOICE3_CAMPPLUS_PATH
• CRISPASR_COSYVOICE3_CFG_BATCH
• CRISPASR_COSYVOICE3_CFG_INTERVAL
• CRISPASR_COSYVOICE3_CFG_INTERVAL_DEBUG
• CRISPASR_COSYVOICE3_DUMP_HIFT
• CRISPASR_COSYVOICE3_DUMP_MEL
• CRISPASR_COSYVOICE3_DUMP_TOKENS
• CRISPASR_COSYVOICE3_FASTCONV
• CRISPASR_COSYVOICE3_FASTCONV_DEBUG
• CRISPASR_COSYVOICE3_SIMDCONV — CPU-only direct SIMD Conv1d path for the 72 HiFT ResBlock convolutions; default ON since the Kaggle quiet-box A/B (1.07x on Xeon avx512f, 1.34x on Zen 4, output 1-LSB-equal, roundtrip exact). Set =0 for the ggml path.
• CRISPASR_COSYVOICE3_SIMDCONV_DEBUG — print pack count, selected ISA, and GPU fallback status.
• CRISPASR_COSYVOICE3_FLOW_STEPS
• CRISPASR_COSYVOICE3_FORCE_GALLOCR
• CRISPASR_COSYVOICE3_GREEDY
• CRISPASR_COSYVOICE3_HIFT_ON_GPU
• CRISPASR_COSYVOICE3_HIFT_PATH
• CRISPASR_COSYVOICE3_KV_BUCKET
• CRISPASR_COSYVOICE3_NO_CLONE_CACHE — re-extract the --voice ref.wav speaker (s3tokenizer + CAMPPlus + prompt mel) on every synthesis instead of once per reference. Output-identical; the cached path is ~30% faster on a multi-sentence --tts (#334).
• CRISPASR_COSYVOICE3_NO_MIN_LEN — drop the decode’s minimum-length floor (2 speech tokens per target text token, upstream’s min_token_text_ratio). Without the floor a single unlucky sample at step 0 ends the decode with no audio at all (#334).
• CRISPASR_COSYVOICE3_UPSTREAM_DIR
• CRISPASR_COSYVOICE3_VOICES_PATH
• CRISPASR_COSYVOICE3_VULKAN_NATIVE

CosyVoice3 (diff-harness assets)

• CRISPASR_CV3_CAMPPLUS_GGUF
• CRISPASR_CV3_FLOW_GGUF
• CRISPASR_CV3_HIFT_GGUF
• CRISPASR_CV3_S3TOK_GGUF

CREPE (pitch)

• CRISPASR_CREPE_BATCH
• CRISPASR_CREPE_DEBUG
• CRISPASR_CREPE_NO_BAKE_F32
• CRISPASR_CREPE_NO_GPU

CSM TTS

• CRISPASR_CSM_BENCH
• CRISPASR_CSM_WAV_FRAMES
• CRISPASR_CSM_WAV_OUT
• CRISPASR_CSM_WAV_TEMP
• CRISPASR_CSM_WAV_TEXT

Dia TTS

• CRISPASR_DIA_BENCH
• CRISPASR_DIA_DECODE_CODES
• CRISPASR_DIA_DUMP_DIR
• CRISPASR_DIA_DUMP_STEPLOGITS
• CRISPASR_DIA_DUMP_TOKENS
• CRISPASR_DIA_FORCE_TOKENS
• CRISPASR_DIA_GREEDY
• CRISPASR_DIA_MAX_STEPS
• CRISPASR_DIA_TTS_GPU

dots.tts

• CRISPASR_DOTS_CFG_INTERVAL
• CRISPASR_DOTS_CFG_INTERVAL_DEBUG
• CRISPASR_DOTS_DIFF_GPU
• CRISPASR_DOTS_DIT_DEBUG
• CRISPASR_DOTS_EOS_THRESHOLD
• CRISPASR_DOTS_FAST
• CRISPASR_DOTS_FM_AB
• CRISPASR_DOTS_FM_DUMP
• CRISPASR_DOTS_FUSED_STEP
• CRISPASR_DOTS_MAX_PATCHES
• CRISPASR_DOTS_ODE_STEPS
• CRISPASR_DOTS_PENC_VERIFY
• CRISPASR_DOTS_TTS_BENCH
• CRISPASR_DOTS_TTS_CPU
• CRISPASR_DOTS_TTS_DEBUG

ECAPA (LID / speaker)

• CRISPASR_ECAPA_ASP_CPU
• CRISPASR_ECAPA_ASP_GGML
• CRISPASR_ECAPA_FORCE_SCALAR
• CRISPASR_ECAPA_LID_BENCH
• CRISPASR_ECAPA_REF_FBANK
• CRISPASR_ECAPA_TIMING

F5-TTS

• CRISPASR_F5_BATCH_CFG
• CRISPASR_F5_BENCH
• CRISPASR_F5_CFG_INTERVAL
• CRISPASR_F5_DIT_SKIP
• CRISPASR_F5_DURATION_CLAMP — clamp the per-char speech rate into a sane English band so a reference whose audio/transcript lengths are mismatched can’t truncate (or balloon) the output (#294). Default on; set 0 to restore the exact upstream ref_T / ref_text_len * gen_text_len / speed estimate.
• CRISPASR_F5_EMBED_GPU
• CRISPASR_F5_F16_ACT
• CRISPASR_F5_HIFIGAN_CPU — 1 restores the pre-a72fb66d CPU-loop HiFi-GAN decode (A/B fallback; the default ggml core_hifigan graph path is ~250x faster on GPU and cosine-1.000000 identical).
• CRISPASR_F5_VOCODE_MEL — debug probe: vocode this mel dump directly, bypassing the DiT (used for the CPU-vs-graph vocoder parity A/B).
• CRISPASR_F5_FORCE_SCALAR
• CRISPASR_F5_REF_MAX_SEC — clip the reference audio to this many seconds before it drives the duration estimate (upstream parity: 12 s). Default 12; set 0 to disable the clip.
• CRISPASR_F5_REF_TRIM_SILENCE — strip leading/trailing silence and collapse internal silences >~1 s in the reference audio (upstream parity). Default on; set 0 to disable.
• CRISPASR_F5_TEXT_LEN_BYTES

FastConformer (shared encoder)

• CRISPASR_FC_ATTN_CONT
• CRISPASR_FC_BUCKET
• CRISPASR_FC_FUSED_QKV
• CRISPASR_FC_GPU_MANUAL_ATTN
• CRISPASR_FC_MAX_LAYERS
• CRISPASR_FC_MEM_DEBUG
• CRISPASR_FC_NO_FLASH
• CRISPASR_FC_PROFILE
• CRISPASR_FC_PROF_FP
• CRISPASR_FC_PROF_FP_COLS
• CRISPASR_FC_PW_Q8
• CRISPASR_FC_TILED_ATTN
• CRISPASR_FC_TILED_BLOCK
• CRISPASR_FC_WINDOWED_ATTN
• CRISPASR_FC_WINDOW_BLOCK

FastPitch

• CRISPASR_FASTPITCH_BENCH
• CRISPASR_FASTPITCH_DUMP_DIR
• CRISPASR_FASTPITCH_FORCE_TOKENS

Ffmpeg

• CRISPASR_FFMPEG_LOG

FireRed ASR / VAD

• CRISPASR_FIRERED_BEAM_F32
• CRISPASR_FIRERED_BENCH
• CRISPASR_FIRERED_ENC_CPU
• CRISPASR_FIRERED_GGML_ATTN
• CRISPASR_FIRERED_LID_BENCH
• CRISPASR_FIRERED_MATVEC_CACHE
• CRISPASR_FIRERED_NO_REPEAT_BREAK
• CRISPASR_FIRERED_VAD_BENCH
• CRISPASR_FIRERED_VAD_DEBUG
• CRISPASR_FIRERED_VAD_FORCE_SCALAR
• CRISPASR_FIRERED_VAD_SERIAL

FunASR / SenseVoice

• CRISPASR_FUNASR_BENCH
• CRISPASR_FUNASR_DUMP_STAGES
• CRISPASR_FUNASR_EMBED_FAST
• CRISPASR_FUNASR_LLM_CPU
• CRISPASR_FUNASR_LLM_LAYERS
• CRISPASR_FUNASR_NAN_CHECK
• CRISPASR_FUNASR_NO_FA
• CRISPASR_FUNASR_ENC_CACHE — 0 disables the exact-T_lfr encoder graph cache (repeat-length calls skip the 7.5–23.9 ms graph rebuild). Default on.
• CRISPASR_FUNASR_STEP_BUCKET — width of the cached decode-graph Lk buckets (default 16, the measured optimum; >= kv_max_ctx reproduces the old fixed-Lk design, which is a ~69% decode regression — A/B arm only).
• CRISPASR_FUNASR_STEP_CACHE — 0 disables the bucketed per-step decode graph cache (bit-identical either way). Default on.

Gemma-4 E2B

• CRISPASR_GEMMA4_AUTO_CHUNK
• CRISPASR_GEMMA4_E2B_BENCH
• CRISPASR_GEMMA4_E2B_EMBED_FAST
• CRISPASR_NO_REL_POS — drop the relative-position (matrix BD) term from the Gemma-4 E2B encoder attention scores, leaving only the content term. Parity bisection aid; not prefixed per-backend for historical reasons.

GLM-ASR

• CRISPASR_GLM_ASR_BENCH
• CRISPASR_GLM_ASR_DEBUG
• CRISPASR_GLM_ASR_EMBED_FAST
• CRISPASR_GLM_ASR_LEGACY_PROMPT
• CRISPASR_GLM_ASR_SINGLE_WINDOW

Granite speech / NLE

• CRISPASR_GRANITE_BENCH
• CRISPASR_GRANITE_DEC_GALLOCR
• CRISPASR_GRANITE_DEC_PROFILE
• CRISPASR_GRANITE_DISABLE_ENCODER_GRAPH
• CRISPASR_GRANITE_ENC_F16
• CRISPASR_GRANITE_FORCE_SCALAR
• CRISPASR_GRANITE_NLE_BENCH
• CRISPASR_GRANITE_NLE_EDIT_DUMP
• CRISPASR_GRANITE_QUANT_ALL

HiFT vocoder

• CRISPASR_HIFT_FULL_IDFT

HTDemucs (source separation)

• CRISPASR_HTDEMUCS_BLAS — route the CrossTransformer matmuls through cblas_sgemm (default ON where Accelerate is available). =0 selects the scalar path. The transformer is ~86% of an unoptimised forward pass, so this is the dominant knob (measured 44x on the transformer, 4.6x overall).
• CRISPASR_HTDEMUCS_FASTCONV — batched im2col + one GEMM for the CPU convs (default ON). =0 selects the original per-time-frame scalar path. Measured enc.conv2d 10.0 s -> 0.17 s and enc.rewrite 12.2 s -> 0.30 s.
• CRISPASR_HTDEMUCS_WCACHE — cache F32 copies of weight tensors by pointer (default ON). =0 re-reads and re-converts on every access, which the DConv stacks do ~6k times per encoder layer.
• CRISPASR_HTDEMUCS_GGML — run the ggml graph path instead of CPU/BLAS. Since #414 the default is AUTO: ON exactly when a real GPU backend is present and permitted (where the fused graph measured ~20x faster than BLAS — RTF 0.37 vs 7.4 on an RTX 3090 Ti), OFF on CPU-only hosts (where graphs measured slower than BLAS). =1/=0 force either way.
• CRISPASR_HTDEMUCS_GPU — GPU permission (CUDA > Metal > Vulkan). Default AUTO follows the caller’s use_gpu (CLI default on); an explicit =0/=1 beats the caller in both directions — so =0 genuinely opts out even though the CLI defaults use_gpu=true (#414 review catch). On GPU-less hosts everything resolves to the BLAS path regardless.
• CRISPASR_HTDEMUCS_NO_BCAST_CAST — disable the issue-#398 fix that casts non-F32 affine/bias weights to F32 in-graph before broadcast add/mul sites (bisection aid). With =1 the pre-fix graph is rebuilt, which on CUDA aborts in binbcast.cu (nb10 % sizeof(src1_t)) because the F16 GGUF stores the DConv GroupNorm affines (*.dconv.layers.N.4.weight) as F16.
• CRISPASR_HTDEMUCS_PROFILE — print a per-phase wall-time breakdown of one forward pass (stft / enc / transformer / dec / istft).
• CRISPASR_HTDEMUCS_DEBUG — verbose per-layer shape and NaN diagnostics.
• CRISPASR_HTDEMUCS_SKIP_TIME — skip the time branch (bisection aid).
• CRISPASR_HTDEMUCS_FUSED — single fused graph (encoder+transformer+decoder on-device, no per-layer host↔device roundtrips). Default AUTO: ON with the GPU graph path (#414), OFF otherwise. =1 alone implies the graph path it needs; =0 on GPU keeps the per-layer-graph bisection arm. The full decision table is unit-locked in tests/test-htdemucs-gates.cpp.
• CRISPASR_HTDEMUCS_MEMSTATS — log each weight-cache admission and the running cache total in MB.
• CRISPASR_HTDEMUCS_NO_SEGMENT — process the whole track in one pass instead of the 25%-overlap segment schedule (A/B against the old behaviour).

All three optimisation gates are output-equivalent: the per-stage diff reports 45/45 stages passing with them ON or OFF.

Higgs STT

• CRISPASR_HIGGS_DEBUG
• CRISPASR_HIGGS_STT_BENCH
• CRISPASR_HIGGS_STT_EMBED_FAST
• CRISPASR_HIGGS_STT_FUSED_QKV

IndexTTS

• CRISPASR_INDEXTTS_AA_BACKEND
• CRISPASR_INDEXTTS_AA_SCALAR
• CRISPASR_INDEXTTS_AUDIO24K_FILE
• CRISPASR_INDEXTTS_BEAM_SIZE
• CRISPASR_INDEXTTS_BENCH
• CRISPASR_INDEXTTS_COND_FILE
• CRISPASR_INDEXTTS_DEBUG
• CRISPASR_INDEXTTS_KV_DEVICE_COPY
• CRISPASR_INDEXTTS_LATENT_FILE
• CRISPASR_INDEXTTS_MEL_CODES_FILE
• CRISPASR_INDEXTTS_MEL_FILE
• CRISPASR_INDEXTTS_SPK_NORM
• CRISPASR_INDEXTTS_TEXT_NORMALIZER
• CRISPASR_INDEXTTS_VOCODER_AA
• CRISPASR_INDEXTTS_VOCODER_RAW
• CRISPASR_INDEXTTS_VOC_BENCH
• CRISPASR_INDEXTTS_VOC_FORCE_GPU

Irodori TTS

• CRISPASR_IRODORI_CAPTION
• CRISPASR_IRODORI_CAPTION_TOKEN_IDS
• CRISPASR_IRODORI_CFG_CAPTION
• CRISPASR_IRODORI_CFG_INTERVAL
• CRISPASR_IRODORI_CFG_INTERVAL_DEBUG
• CRISPASR_IRODORI_CFG_SPEAKER
• CRISPASR_IRODORI_CFG_TEXT
• CRISPASR_IRODORI_CODEC_CPU
• CRISPASR_IRODORI_CODEC_GPU
• CRISPASR_IRODORI_CPU
• CRISPASR_IRODORI_DEBUG
• CRISPASR_IRODORI_DECODE_CHUNK
• CRISPASR_IRODORI_DECODE_CTX
• CRISPASR_IRODORI_DIT_TIMING
• CRISPASR_IRODORI_DUMP_LATENT
• CRISPASR_IRODORI_DUMP_TEXT_STATE
• CRISPASR_IRODORI_DUMP_TOKENS
• CRISPASR_IRODORI_DUMP_V_PRED0
• CRISPASR_IRODORI_ENC_DUMP
• CRISPASR_IRODORI_ENC_PRENORM
• CRISPASR_IRODORI_FASTCONV
• CRISPASR_IRODORI_LAYERS
• CRISPASR_IRODORI_ODE_STEPS
• CRISPASR_IRODORI_PERSIST_GRAPH
• CRISPASR_IRODORI_REF_NOISE
• CRISPASR_IRODORI_TOKEN_IDS
• CRISPASR_IRODORI_T_LATENT

Kokoro

• CRISPASR_KOKORO_BENCH
• CRISPASR_KOKORO_DEBUG
• CRISPASR_KOKORO_DEBUG_INTERMEDIATES
• CRISPASR_KOKORO_DUMP_STAGES
• CRISPASR_KOKORO_FASTCONV
• CRISPASR_KOKORO_FASTCONV_DEBUG
• CRISPASR_KOKORO_G2P
• CRISPASR_KOKORO_GEN_FORCE_METAL / CRISPASR_KOKORO_GEN_GPU — either one puts the generator (vocoder) stage on Metal.
• CRISPASR_KOKORO_SEED
• CRISPASR_KOKORO_USE_GPU
• CRISPASR_KOKORO_VOICE_GGUF

KugelAudio

• CRISPASR_KUGELAUDIO_CPU_ONLY
• CRISPASR_KUGELAUDIO_DEBUG

Kyutai STT

• CRISPASR_KYUTAI_RVQ_FAST
• CRISPASR_KYUTAI_STT_BENCH

LFM2-Audio

• CRISPASR_LFM2_AUDIO_BENCH
• CRISPASR_LFM2_AUDIO_CPU
• CRISPASR_LFM2_SNAP_LAYERS

M2M-100 translate

• CRISPASR_M2M100_BENCH
• CRISPASR_M2M100_GPU

MarbleNet VAD

• CRISPASR_MARBLENET_VAD_BENCH
• CRISPASR_MARBLENET_VAD_SERIAL

Mel-Band RoFormer (source separation)

• CRISPASR_MBR_PROFILE — print a per-stage wall-time breakdown of one forward pass (stft+pack / band_split / run_time / run_freq / mask_est / synthesize).
• CRISPASR_MELBAND_GGML — run the ggml graph path instead of the legacy CPU path. Since the Change-176 graph port the default is AUTO: ON exactly when a real GPU backend is present and permitted (the fused single graph measured ~112x faster than the per-layer graphs — RTF ~0.09 vs ~10 on an RTX 3090 Ti), OFF on CPU-only hosts. =1/=0 force either way.
• CRISPASR_MELBAND_GPU — GPU permission (CUDA > Metal > Vulkan). Default AUTO follows the caller’s use_gpu (CLI default on); an explicit =0/=1 beats the caller in both directions — so =0 genuinely opts out even though the CLI defaults use_gpu=true (#414 review semantics). On GPU-less hosts everything resolves to the CPU path regardless.
• CRISPASR_MELBAND_FUSED — single fused graph: band-split + the full time/freq transformer stack + mask estimator on-device in one graph, no per-layer host↔device roundtrips (the measured-fastest path on GPU). Default AUTO: ON with the GPU graph path, OFF otherwise. =1 alone implies the graph path it needs; =0 on GPU keeps the per-layer-graph bisection arm. The full decision table is unit-locked in tests/test-mel-band-gates.cpp.
• CRISPASR_MELBAND_SEG_S — override the Demucs-style segment length in seconds (params.segment_seconds; <=0 → the checkpoint’s trained chunk_size from GGUF metadata, 8 s Kim fallback — do NOT expect 10 s: the earlier hardcoded 10 s ran RoPE 25% past the trained window, review #422). The attention matrix is O(T²·bands·heads), so long inputs are split into segments with 25% overlap and a triangular weight (bounds VRAM; the unsegmented whole-buffer path OOMs on any clip beyond ~10 s).
• CRISPASR_MELBAND_NO_SEGMENT — process the whole track in one pass instead of the segmented overlap-add schedule (A/B against the old behaviour).

MeloTTS

• CRISPASR_MELOTTS_BENCH
• CRISPASR_MELOTTS_BERT
• CRISPASR_MELOTTS_FORCE_SCALAR
• CRISPASR_MELOTTS_WEIGHT_CACHE

Mimi codec

• CRISPASR_MIMI_NONCAUSAL

MiMo-ASR

• CRISPASR_MIMO_ASR_BENCH
• CRISPASR_MIMO_ASR_DIAG
• CRISPASR_MIMO_ASR_DUMP_STAGES
• CRISPASR_MIMO_ASR_GPU
• CRISPASR_MIMO_FORCE_CPU
• CRISPASR_MIMO_SMOKE_DUMP
• CRISPASR_MIMO_SMOKE_GPU
• CRISPASR_MIMO_TOKENIZER_GPU
• CRISPASR_MIMO_TOK_CPU
• CRISPASR_MIMO_TOK_CPU_RVQ
• CRISPASR_MIMO_TOK_VERIFY_RVQ

mini-omni2

• CRISPASR_MINI_OMNI2_BENCH

Moonshine

• CRISPASR_MOONSHINE_ALL_GPU
• CRISPASR_MOONSHINE_BENCH
• CRISPASR_MOONSHINE_ENC_ATTN
• CRISPASR_MOONSHINE_NO_REPEAT_BREAK
• CRISPASR_MOONSHINE_STREAMING_BENCH
• CRISPASR_MOONSHINE_STREAMING_GPU
• CRISPASR_MOONSHINE_STREAM_BENCH

MOSS family

• CRISPASR_MOSS_AUDIO_BENCH
• CRISPASR_MOSS_AUDIO_EMBED_FAST
• CRISPASR_MOSS_AUDIO_ENC_FLASH
• CRISPASR_MOSS_AUDIO_ENC_MANUAL
• CRISPASR_MOSS_AUDIO_FORCE_CPU
• CRISPASR_MOSS_AUDIO_MEL_FILE
• CRISPASR_MOSS_DIARIZE_BENCH
• CRISPASR_MOSS_DIARIZE_DEBUG
• CRISPASR_MOSS_DIARIZE_DUMP_CONV
• CRISPASR_MOSS_DIARIZE_ENC_FLASH
• CRISPASR_MOSS_DIARIZE_ENC_MANUAL
• CRISPASR_MOSS_DIARIZE_FORCE_CPU
• CRISPASR_MOSS_DIARIZE_NO_LOOPFIX
• CRISPASR_MOSS_TRANSCRIBE_BENCH
• CRISPASR_MOSS_TRANSCRIBE_ENC_DUMP
• CRISPASR_MOSS_TRANSCRIBE_ENC_FLASH
• CRISPASR_MOSS_TRANSCRIBE_ENC_MANUAL
• CRISPASR_MOSS_TRANSCRIBE_FORCE_CPU
• CRISPASR_MOSS_TRANSCRIBE_L0_DUMP
• CRISPASR_MOSS_TRANSCRIBE_MEL_DUMP
• CRISPASR_MOSS_TRANSCRIBE_NO_LOOPFIX
• CRISPASR_MOSS_TTS_BENCH
• CRISPASR_MOSS_TTS_LOCAL_DEBUG
• CRISPASR_MOSS_TTS_LOCAL_DUMP_FA_PATH
• CRISPASR_MOSS_TTS_LOCAL_DUMP_HIDDEN
• CRISPASR_MOSS_TTS_LOCAL_DUMP_LAYERS
• CRISPASR_MOSS_TTS_LOCAL_DUMP_PROMPT_IDS
• CRISPASR_MOSS_TTS_LOCAL_DUMP_STOP
• CRISPASR_MOSS_TTS_LOCAL_DUMP_SUBLAYER
• CRISPASR_MOSS_TTS_LOCAL_DUMP_SUBLAYER_PATH
• CRISPASR_MOSS_TTS_LOCAL_FORCE_FRAMES
• CRISPASR_MOSS_TTS_LOCAL_GREEDY_AUDIO
• CRISPASR_MOSS_TTS_LOCAL_GREEDY_TEXT
• CRISPASR_MOSS_TTS_LOCAL_INJECT_LAYER
• CRISPASR_MOSS_TTS_LOCAL_INJECT_PATH
• CRISPASR_MOSS_TTS_LOCAL_MAX_FRAMES
• CRISPASR_MOSS_TTS_LOCAL_NO_GPU

MP3 codec

• CRISPASR_MP3_ENCODER

Diarization — foxnose (#324)

• CRISPASR_DIARIZE_COUNT — speaker-count estimator: bic (default, the upstream GMM/BIC + silhouette sweep) or eigengap. Eigengap is better on well-separated synthetic data and cheaper, but under-counts on real speech (11.4 % vs 5.3 % DER on VoxConverse) — see docs/foxnose-diarize/PLAN.md
• CRISPASR_DIARIZE_BIC_WINDOW — score silhouette only in a [k-2, k+3] window around the BIC anchor instead of the full [min, max] range. The full range is the default: the BIC anchor is unreliable in both directions (measured errors of +5 / -3 / -3 on 4/5/6 well-separated blobs) and when it over-counts the window is stranded above the truth and cannot climb back to it
• CRISPASR_WESPEAKER_BENCH — per-stage embedder timings (fbank / resnet / resnet_windows). Counting invocations of these is also how you check WHICH embedding path actually ran
• CRISPASR_WESPEAKER_DEBUG — embedder diagnostics
• CRISPASR_DIARIZE_DEBUG — chosen speaker count, the reason it was chosen, and the per-k silhouette curve behind it. Worth reading before trusting a count: on a borderline file the decision can rest on a <1 % score gap
• CRISPASR_DIARIZE_EMBED_WORKERS — windows embedded concurrently (default: -t). Each worker gets its own context sharing one copy of the weights
• CRISPASR_SPEAKER_EMBED_WORKERS — the same worker count for the standalone crispasr-diarize CLI’s speaker-embedding stage
• CRISPASR_SPEAKER_EMBED_THREADS — ggml threads per embedder context (default: -t). Honoured by the pluggable embedders and by wespeaker
• CRISPASR_DIARIZE_SPAN_EMBED=1 — run ONE network pass per span of windows instead of one per window. 1.78x less diarization CPU for +0.30 mean DER on the VoxConverse dev shard; off by default because accuracy is the better default for a diarizer. See docs/cli.md#diarization
• CRISPASR_DIARIZE_SPAN_WINDOWS — windows per span (default 32). Measured NOT to affect the accuracy cost — identical from N=2 to N=32 — so there is nothing to tune here; larger is simply faster
• CRISPASR_WESPEAKER_CONV — conv lowering for the WeSpeaker embedder: im2col (default) lowers each conv to explicit IM2COL + MUL_MAT nodes so the GEMMs reach the Accelerate BLAS backend on CPU and the simdgroup mul_mm kernels on Metal; direct restores GGML_OP_CONV_2D. Measured on esrit.wav (215 s, -t 8): diarization delta 9.8 s -> 5.9 s (~1.6x), embeddings cosine 1.0 vs direct, DER identical per file (7.32 % shard mean)
• CRISPASR_WESPEAKER_GPU=1 — run the WeSpeaker embedder on the GPU backend (single context; the CPU worker pool is stood down because workers borrow weights that now live in a GPU buffer). With the im2col default and batched windows Metal reaches parity with the 8-worker CPU schedule on an M-series (8.5 s vs 8.7 s wall on esrit.wav) but does not beat it, so CPU stays the default; the switch exists for machines where the GPU/CPU balance differs
• CRISPASR_DIARIZE_BATCH_EMBED — batch independent 1.2 s windows into one graph along ne[3] (arithmetic-identical to per-window; cosine 1.0). Default: ON under CRISPASR_WESPEAKER_GPU=1 (collapses ~350 Metal dispatches into ~11 and is what got Metal from a 2x loss to parity), OFF on CPU (measured 12.1 -> 14.9 s wall on esrit.wav: ggml’s CPU conv loops ne[3], so fusing buys no GEMM shape and the 32-window chunks starve the worker schedule). 1/0 forces either way
• CRISPASR_WESPEAKER_BATCH — max windows per batched graph (default 32 on GPU, 16 on CPU; cap 32). Only meaningful where the batch path is active

GigaAM-v3

• CRISPASR_GIGAAM_BENCH — per-stage timings (mel / encoder / decode)
• CRISPASR_GIGAAM_DEBUG — encoder output min/max/mean
• CRISPASR_GIGAAM_FLASH — ggml_flash_attn_ext in the encoder (opt-in; the manual QK^T path is what the per-stage diff was validated on)
• CRISPASR_GIGAAM_FORCE_SCALAR — scalar LSTM/joint loops instead of cblas
• CRISPASR_GIGAAM_QUANT_ALL — let crispasr-quantize quantize the heads and the pre-encode convs too (default keeps them at source precision)

Nemotron

• CRISPASR_NEMOTRON_BENCH
• CRISPASR_NEMOTRON_CONTEXT_PRESET
• CRISPASR_NEMOTRON_DEBUG
• CRISPASR_NEMOTRON_DECODE_TIMING
• CRISPASR_NEMOTRON_FORCE_SCALAR
• CRISPASR_NEMOTRON_GGML_DECODE
• CRISPASR_NEMOTRON_MAES
• CRISPASR_NEMOTRON_NO_WINDOW_MASK
• CRISPASR_NEMOTRON_STREAMING
• CRISPASR_NEMOTRON_STREAM_DEBUG

OmniASR

• CRISPASR_OMNIASR_BENCH
• CRISPASR_OMNIASR_DEBUG
• CRISPASR_OMNIASR_DUMP_DIR
• CRISPASR_OMNIASR_KEEP_F16_HEAD
• CRISPASR_OMNIASR_KEEP_F16_TAIL
• CRISPASR_OMNIASR_QUANT_ALL

OmniVoice

• CRISPASR_OMNIVOICE_ACENC_BISECT
• CRISPASR_OMNIVOICE_AUTO_LANG — default on. When no language was requested (-l / -tl / the server’s "language" / set_target_language), guess one from the text being spoken and use it if it maps to an id the model knows. An explicitly requested language always wins; a low-confidence guess resolves to nothing and behaves exactly as before. =0 restores the old always-language-agnostic behaviour. Exists because SubtitleEdit’s language menu is not yet wired to its request payload (#13273).
• CRISPASR_OMNIVOICE_BENCH
• CRISPASR_OMNIVOICE_CFG_INTERVAL
• CRISPASR_OMNIVOICE_CHUNK
• CRISPASR_OMNIVOICE_CLASS_TEMP
• CRISPASR_OMNIVOICE_CODEC_FASTCONV
• CRISPASR_OMNIVOICE_CODEC_GPU — codec placement override (1 = GPU, 0 = CPU). Unset defaults to GPU on CUDA and CPU on Metal/CPU.
• CRISPASR_OMNIVOICE_CPU
• CRISPASR_OMNIVOICE_DEBUG
• CRISPASR_OMNIVOICE_DEBUG_CODES
• CRISPASR_OMNIVOICE_DEBUG_SUM
• CRISPASR_OMNIVOICE_DUMP_CODES
• CRISPASR_OMNIVOICE_ENCODE_DIFF
• CRISPASR_OMNIVOICE_FRAMES_PER_CHAR
• CRISPASR_OMNIVOICE_FUSED_STEP
• CRISPASR_OMNIVOICE_GUIDANCE
• CRISPASR_OMNIVOICE_HUBERT_REF
• CRISPASR_OMNIVOICE_NUM_STEPS
• CRISPASR_OMNIVOICE_PERSISTENT_GRAPH
• CRISPASR_OMNIVOICE_POS_TEMP
• CRISPASR_OMNIVOICE_REF_RATE_CHECK
• CRISPASR_OMNIVOICE_TOKENIZER_GGUF
• CRISPASR_OMNIVOICE_UNIFIED_CFG
• CRISPASR_OMNIVOICE_UPSTREAM_WEIGHTS
• CRISPASR_OMNIVOICE_VOICE_CACHE

OpenVoice2

• CRISPASR_OPENVOICE2_BENCH
• CRISPASR_OV2_DUMP_DIR
• CRISPASR_OV2_FORCE_SCALAR
• CRISPASR_OV2_NO_NORMALIZE
• CRISPASR_OV2_TAU

Opus codec

• CRISPASR_OPUS_DEBUG
• CRISPASR_OPUS_DECODER
• CRISPASR_OPUS_ENCODER

Orpheus

• CRISPASR_ORPHEUS_BENCH
• CRISPASR_ORPHEUS_BUCKET
• CRISPASR_ORPHEUS_DEBUG
• CRISPASR_ORPHEUS_DIFF_GPU
• CRISPASR_ORPHEUS_DIFF_MAXGEN
• CRISPASR_ORPHEUS_PROMPT_IDS
• CRISPASR_ORPHEUS_SNAC_CODE
• CRISPASR_ORPHEUS_SNAC_GPU
• CRISPASR_ORPHEUS_SNAC_T_SUPER

OuteTTS

• CRISPASR_OUTETTS_BENCH

Paraformer

• CRISPASR_PARAFORMER_BENCH
• CRISPASR_PARAFORMER_GPU

Parakeet

• CRISPASR_PARAKEET_ATT_CONTEXT
• CRISPASR_PARAKEET_BENCH
• CRISPASR_PARAKEET_CHUNK_OVERLAP
• CRISPASR_PARAKEET_CHUNK_SECONDS
• CRISPASR_PARAKEET_DEBUG
• CRISPASR_PARAKEET_DECODE_TIMING
• CRISPASR_PARAKEET_ENC_CACHE
• CRISPASR_PARAKEET_ENC_PROBE
• CRISPASR_PARAKEET_FORCE_SCALAR
• CRISPASR_PARAKEET_GGML_DECODE
• CRISPASR_PARAKEET_INTERNAL_CHUNKING
• CRISPASR_PARAKEET_LONGFORM
• CRISPASR_PARAKEET_LONGFORM_WINDOW
• CRISPASR_PARAKEET_MAES
• CRISPASR_PARAKEET_MEM_COEFF
• CRISPASR_PARAKEET_MEM_POLICY
• CRISPASR_PARAKEET_PIPELINE
• CRISPASR_PARAKEET_QUANT_ALL
• CRISPASR_PARAKEET_SIMULATE_ENCODE_OOM
• CRISPASR_PARAKEET_STREAM_CHUNK
• CRISPASR_PARAKEET_STREAM_OVERLAP
• CRISPASR_PARAKEET_STREAM_THRESHOLD
• CRISPASR_PARAKEET_VAD_SLICE_CAP
• CRISPASR_PARAKEET_VRAM_BUDGET_MB

Parler-TTS

• CRISPASR_PARLER_BUCKET
• CRISPASR_PARLER_DEBUG
• CRISPASR_PARLER_DESC_IDS
• CRISPASR_PARLER_DIFF_MAXGEN
• CRISPASR_PARLER_DUMP_ENC
• CRISPASR_PARLER_PROMPT_IDS
• CRISPASR_PARLER_TTS_BENCH

Piano transcription

• CRISPASR_PIANO_SERIAL — 1 restores the single-threaded conv/GRU/linear compute (the default parallel path is bit-identical; #305, ~1.7x).

Piper

• CRISPASR_PIPER_FORCE_SCALAR
• CRISPASR_PIPER_TTS_BENCH
• CRISPASR_PIPER_WEIGHT_CACHE

Pocket-TTS

• CRISPASR_POCKET_DUMP_DIR
• CRISPASR_POCKET_FORCE_LATENTS
• CRISPASR_POCKET_FORCE_NOISE
• CRISPASR_POCKET_MANUAL_BACKBONE
• CRISPASR_POCKET_MANUAL_MIMI
• CRISPASR_POCKET_MAX_FRAMES
• CRISPASR_POCKET_MIMI_DUMP
• CRISPASR_POCKET_MIMI_SCALAR
• CRISPASR_POCKET_TTS_BENCH
• CRISPASR_POCKET_VOICE_CACHE
• CRISPASR_POCKET_VULKAN_MIMI_MAX_FRAMES

Pyannote segmentation

• CRISPASR_PYANNOTE_LEGACY
• CRISPASR_PYANNOTE_SEG_BENCH
• CRISPASR_PYANNOTE_SEG_DUMP
• CRISPASR_PYANNOTE_CHUNK_S — audio per chunk of parallel inference, in seconds (default 60; 0 restores the pre-#326 single scan over the whole file). Chunking is decided by audio LENGTH and never by thread count, so posteriors do not change with -t
• CRISPASR_PYANNOTE_CHUNK_CONTEXT_S — real audio spliced either side of a chunk and then trimmed (default 5), which absorbs the convolutions’ zero padding and the LSTM’s zero initial state

Qwen3-ASR

• CRISPASR_QWEN3ASR_QUANT_AUDIO

Qwen3-ASR / Qwen3-TTS

• CRISPASR_QWEN3_ASR_BENCH
• CRISPASR_QWEN3_ASR_EMBED_FAST
• CRISPASR_QWEN3_ASR_FUSED_QKV
• CRISPASR_QWEN3_SYSPROMPT_LANG
• CRISPASR_QWEN3_TTS_BENCH
• CRISPASR_QWEN3_TTS_CODEC_ALLOW_FULL
• CRISPASR_QWEN3_TTS_CODEC_CHUNK
• CRISPASR_QWEN3_TTS_CODEC_CPU
• CRISPASR_QWEN3_TTS_CODEC_CTX
• CRISPASR_QWEN3_TTS_CODEC_FASTCONV — codec conv fast path. Default ON; 0 opts out.
• CRISPASR_QWEN3_TTS_CODEC_FORCE_METAL
• CRISPASR_QWEN3_TTS_CODEC_GGUF
• CRISPASR_QWEN3_TTS_CODEC_GPU
• CRISPASR_QWEN3_TTS_HIP_CODEC_NATIVE — bypass the #337 ROCm codec-encoder correctness fallback and run it natively on HIP. Diagnostic/A/B only until the triggering Daphne spans pass on a real AMD GPU.
• CRISPASR_QWEN3_TTS_HIP_CP_NATIVE — bypass the #337 CPU fallback for the ROCm 0.6B-F16 code predictor. Diagnostic/A/B only; its native gfx1100 path was observed to emit all-NaN logits.
• CRISPASR_QWEN3_TTS_CODEC_TRACE
• CRISPASR_QWEN3_TTS_CP_BACKEND
• CRISPASR_QWEN3_TTS_CP_DIRECT
• CRISPASR_QWEN3_TTS_CP_MTP_NOFUSE
• CRISPASR_QWEN3_TTS_CP_STEP0_CACHE
• CRISPASR_QWEN3_TTS_DEBUG
• CRISPASR_QWEN3_TTS_DUMP_DIR
• CRISPASR_QWEN3_TTS_EMBD_CHECK
• CRISPASR_QWEN3_TTS_FUSED_QKV
• CRISPASR_QWEN3_TTS_LK_BUCKET
• CRISPASR_QWEN3_TTS_NO_EMBD_CACHE
• CRISPASR_QWEN3_TTS_O15 / _O15_SKIP_REALLOC
• CRISPASR_QWEN3_TTS_PROF
• CRISPASR_QWEN3_TTS_SEED
• CRISPASR_QWEN3_TTS_TALKER_SCHED
• CRISPASR_QWEN3_TTS_VULKAN_NATIVE
• CRISPASR_QWEN3_TTS_DUMP_LOGITS=<dir> — write the raw per-frame talker logits (f32, before the repetition penalty and the suppress mask) plus a top-5 line to stderr. The instrument for a cross-backend diff: tokens alone cannot tell a miscompute from amplified rounding (#337).
• CRISPASR_QWEN3_TTS_REPLAY_CODES=<file> — 16 whitespace-separated codec ids per frame; the decode uses them instead of sampling. Teacher forcing, and the ONLY way to compare two backends step by step: pin the whole frame or the 15 residual codebooks (which must be sampled) diverge and the diff measures trajectory, not arithmetic (#337).
• CRISPASR_QWEN3_TTS_REPLAY_TOKENS=<file> — the weaker form: codebook-0 ids only. Useful for forcing a trajectory, NOT sufficient for a logits diff.
• CRISPASR_QWEN3_TTS_GREEDY — force the talker’s codebook-0 sampler to argmax (top_k=1). The frame sequence then depends only on the logits, so two backends agree if and only if their logits agree — this is the lever for telling a GPU miscompute apart from a sampling difference, and without it a CPU-vs-GPU token comparison proves nothing (#337).
• CRISPASR_QWEN3_TTS_MAX_FRAMES
• CRISPASR_QWEN3_TTS_SKIP_REF_DECODE

SenseVoice

• CRISPASR_SENSEVOICE_BENCH
• CRISPASR_SENSEVOICE_NO_FA

Sidon

• CRISPASR_SIDON_FASTCONV — DAC convolution mode (off, k1-f16, k1-f32, or full). Unset defaults to k1-f16 on CUDA and off on Vulkan/CPU.
• CRISPASR_SIDON_RPE — relative-position-bias formulation: bucket-direct, bucket, or expand (legacy [head_dim, T, T] expansion; keeps the Vulkan mul_mat batching branch). All three are algebraically equivalent. Unset defaults to AUTO, resolved per graph build because the choice depends on the input length: expand on a GPU backend while its extra transient footprint (4·T²·(head_dim+1−heads), i.e. 196·T² bytes for the shipped model — 58 MiB at T=557, 1.64 GiB at the 3000-frame cap) fits CRISPASR_SIDON_RPE_BUDGET_MB, and bucket-direct otherwise. AUTO exists because expand measured 2.7× faster in the predictor on the #416 reporter’s GTX 1660 SUPER (213.80 ms vs 575.25 ms at T=557, same file and binary). AUTO stays off on CPU: that speedup is one device’s measurement and CPU behaviour is left exactly as it was. An explicit value is honoured as given and never auto-overridden — it is the #416 bisection handle. The decision table is unit-locked in tests/test-sidon-rpe-gates.cpp.
• CRISPASR_SIDON_RPE_BUDGET_MB — AUTO’s budget in MiB for expand’s extra transient footprint (default 256, which admits expand up to T≈1170, ~23 s of audio). 0 disables AUTO, pinning bucket-direct.
• CRISPASR_SIDON_DECODER_CHUNK_FRAMES — maximum DAC core size in feature frames (default 512). 0 decodes the whole utterance in one graph (~4.5 GiB at T≈2825 vs ~0.79 GiB chunked). Chunked output is bit-exact against the whole-utterance decode.
• CRISPASR_SIDON_LOOKAHEAD — set to 0 to disable the input padding (one leading predictor frame plus 1.5 s of right-side lookahead). Padding is on by default; without it the last ~12 ms of every clip is a full-scale transient.
• CRISPASR_SIDON_MAX_FRAMES — predictor input cap in feature frames (default 3000, ~58.5 s after the lookahead). Guards the O(T^2) attention.
• CRISPASR_SIDON_WINDOW_FRAMES — predictor window size in feature frames. Default 0 = off, i.e. input past CRISPASR_SIDON_MAX_FRAMES is refused as before. Set to e.g. 1500 to process long input as overlapping windows instead (#431). Off by default because the evidence is incomplete: windowed output beat the whole-utterance path on a 50 s ASR roundtrip, but an ASR roundtrip measures “sounds better to a small ASR”, not “faithful to upstream”, and those can disagree. tools/kaggle/sidon-length-parity is the experiment that settles it. The window doubles as the threshold — at T <= window there is one window and the split is a no-op — and it is clamped to CRISPASR_SIDON_MAX_FRAMES so raising one cannot silently violate the other.
• CRISPASR_SIDON_PREDICTOR_CONTEXT_FRAMES — context frames on each side of a window (default 300), cropped away after the run. Unlike the DAC’s chunking, whose cores are exact because the decoder is fully convolutional with a finite receptive field (dac_receptive_frames()), attention has no receptive field, so no context size makes a windowed core exact — this trades compute for boundary quality rather than buying correctness.
• CRISPASR_SIDON_DEBUG — print per-stage scheduler workspace sizes (per backend) after graph allocation.
• CRISPASR_SIDON_DUMP_HANDOFF — directory/path for the predictor→DAC handoff tensor dump.

Sherpa

• CRISPASR_SHERPA_LID_BIN
• CRISPASR_SHERPA_LID_TIMEOUT_SEC / CRISPASR_SHERPA_TIMEOUT_SEC — wall-clock timeout for the external sherpa LID / diarize helper, scaled by audio length.

Silero LID

• CRISPASR_SILERO_FORCE_SCALAR
• CRISPASR_SILERO_LID_BENCH
• CRISPASR_SILERO_LID_DEBUG
• CRISPASR_SILERO_LID_DUMP
• CRISPASR_SILERO_LID_LEGACY
• CRISPASR_SILERO_LID_MAX_S
• CRISPASR_SILERO_LID_MIN_LOGIT — evidence gate (#409): discard a silero LID answer whose top-1 RAW logit is below this floor (default -2.0; -999 disables) and fall back to whisper-tiny LID. The raw-logit magnitude separates in-domain from out-of-domain audio where the softmax probability does not.
• CRISPASR_SILERO_LID_TRACE
• CRISPASR_SILERO_LID_TRACE_OFF
• CRISPASR_SILERO_LID_TRUNC
• CRISPASR_SILERO_LID_VULKAN

SpeechT5

• CRISPASR_SPEECHT5_DUMP_DIR
• CRISPASR_SPEECHT5_FASTCONV
• CRISPASR_SPEECHT5_FASTCONV_DEBUG
• CRISPASR_SPEECHT5_TTS_BENCH

T5 translate

• CRISPASR_T5_DIFF
• CRISPASR_T5_GPU
• CRISPASR_T5_TRANSLATE_BENCH

TabCNN (guitar tablature)

• CRISPASR_TABCNN_DEBUG
• CRISPASR_TABCNN_NO_GPU

TaDa TTS

• CRISPASR_TADA_ACOUSTIC_CFG
• CRISPASR_TADA_ALLOW_VULKAN
• CRISPASR_TADA_BATCH_PREFILL
• CRISPASR_TADA_BENCH
• CRISPASR_TADA_BUCKET_MIN
• CRISPASR_TADA_CFG_INTERVAL
• CRISPASR_TADA_CFG_INTERVAL_DEBUG
• CRISPASR_TADA_CODEC_BENCH
• CRISPASR_TADA_CODEC_DUMP
• CRISPASR_TADA_CODEC_GGUF
• CRISPASR_TADA_CODEC_VULKAN_NATIVE
• CRISPASR_TADA_CTC_ASR
• CRISPASR_TADA_DIFF_TEXT
• CRISPASR_TADA_DO_SAMPLE
• CRISPASR_TADA_DUMP_ACOUSTIC_FEATURES
• CRISPASR_TADA_DUMP_FEATURES
• CRISPASR_TADA_DUMP_FM_STEPS
• CRISPASR_TADA_DUMP_TIME_BEFORE
• CRISPASR_TADA_ENCODER_DEBUG
• CRISPASR_TADA_EXTRA_STEPS
• CRISPASR_TADA_FM_B2
• CRISPASR_TADA_KEEP_F16_HEAD
• CRISPASR_TADA_KEEP_F16_TAIL
• CRISPASR_TADA_MAX_EXPANDED_FRAMES
• CRISPASR_TADA_NOISE_TEMP
• CRISPASR_TADA_NO_BUCKET
• CRISPASR_TADA_NUM_CANDIDATES
• CRISPASR_TADA_NUM_FM_STEPS
• CRISPASR_TADA_PROMPT_CACHE
• CRISPASR_TADA_PROMPT_TEXT
• CRISPASR_TADA_QUANT_ALL
• CRISPASR_TADA_REPETITION_PENALTY
• CRISPASR_TADA_SCORER
• CRISPASR_TADA_TALKER_TIMING
• CRISPASR_TADA_TEMPERATURE
• CRISPASR_TADA_TOP_K
• CRISPASR_TADA_TOP_P
• CRISPASR_TADA_VULKAN_NATIVE
• CRISPASR_TADA_WAV_CLONE — 1 enables on-the-fly TaDa voice cloning from a reference .wav + transcript through the session C-ABI (#201). Off by default: without it a .wav voice reference is still rejected with -2, preserving the historical behaviour until the decoded-output roundtrip has validated the path.

TitaNet speaker

• CRISPASR_TITANET_BENCH
• CRISPASR_TITANET_DUMP
• CRISPASR_TITANET_DUMP_MEL
• CRISPASR_TITANET_FORCE_SCALAR
• CRISPASR_TITANET_GGML
• CRISPASR_TITANET_GPU
• CRISPASR_TITANET_LEGACY
• CRISPASR_TITANET_REF_MEL

Three compute paths exist and all are kept working; the default is the fastest one measured per platform.

path	selected by	measured, M1, 2 s segment
legacy (Accelerate / hand-rolled)	default where HAVE_ACCELERATE	71.7 ms
ggml graph, CPU	default elsewhere; CRISPASR_TITANET_GGML=1	277.3 ms
ggml graph, GPU	CRISPASR_TITANET_GGML=1 CRISPASR_TITANET_GPU=1	31.9 ms (but see below)

All three agree to cosine 1.000000 with each other and 0.999996 against NVIDIA’s nemo_en_titanet_large.onnx export fed the same mel — so the choice is purely about speed.

⚠ CRISPASR_TITANET_GPU=1 is opt-in because it loses on real workloads despite winning the micro-benchmark. Diarization embeds one segment per call at variable lengths, so every call reshapes the graph and the GPU allocator re-reserves; and CRISPASR_SPEAKER_EMBED_WORKERS runs several embedders at once, which contend for the one GPU. End-to-end on a 600 s clip, 47 segments:

workers=4, legacy   9994 ms    <- default, fastest
workers=1, legacy  12673 ms
workers=1, GPU     15275 ms
workers=4, GPU     49866 ms

Keep it for evaluating a discrete GPU (where the balance may differ) or for CRISPASR_SPEAKER_EMBED_WORKERS=1 on a machine with weak CPU cores. Bucketing segment lengths so the graph shape stops changing is the work that would make this path win generally.

CRISPASR_TITANET_DUMP_MEL=<path> writes the computed mel as [T][n_mels] float32 — the counterpart to CRISPASR_TITANET_REF_MEL. Feeding that dump to an upstream ONNX export separates the front-end from the network, which a single end-to-end cosine cannot do.

VibeVoice

• CRISPASR_VIBEVOICE_ASR_PROMPT
• CRISPASR_VIBEVOICE_ASR_SAMPLE — streaming ASR samples the acoustic posterior by default, matching upstream. Set 0 for deterministic posterior-mean reference diffs or 1 to force sampling.
• CRISPASR_VIBEVOICE_ATTN_PREC
• CRISPASR_VIBEVOICE_BENCH
• CRISPASR_VIBEVOICE_BITNET_ACT_QUANT
• CRISPASR_VIBEVOICE_DEBUG
• CRISPASR_VIBEVOICE_DUMP_DIR
• CRISPASR_VIBEVOICE_ENC_BACKEND — {auto|cpu|gpu} backend for the σ-VAE tokenizer ENCODERS (the ASR direction). auto (default) diverts them to CPU on Vulkan devices with a 65535 maxComputeWorkGroupCount (Intel Arc/Iris/ UHD, llvmpipe), where the conv dispatches over long audio overflow and abort (issue #418 — the ASR twin of the decoder’s #52 fallback, CRISPASR_VIBEVOICE_VAE_BACKEND). gpu forces the active backend (the repro arm); cpu forces the fallback anywhere.
• CRISPASR_VIBEVOICE_ENCODER_CHUNK_SECONDS
• CRISPASR_VIBEVOICE_ENCODER_CONTEXT_SECONDS
• CRISPASR_VIBEVOICE_GELU_TANH
• CRISPASR_VIBEVOICE_GPU
• CRISPASR_VIBEVOICE_LM_BUCKETS
• CRISPASR_VIBEVOICE_NO_INPUT_NORM
• CRISPASR_VIBEVOICE_NO_LM_BUCKETS
• CRISPASR_VIBEVOICE_PRED_SCHED
• CRISPASR_VIBEVOICE_QUANT_ALL
• CRISPASR_VIBEVOICE_RAW_TRANSCRIPT — 1 keeps the pre-#300 single segment holding the model’s raw JSON blob instead of splitting it into one segment per utterance with the speaker in the structured field.
• CRISPASR_VIBEVOICE_REF_FEATURES
• CRISPASR_VIBEVOICE_REUSE_PRED_GRAPH
• CRISPASR_VIBEVOICE_TTS_CFG_SCALE
• CRISPASR_VIBEVOICE_TTS_DUMP
• CRISPASR_VIBEVOICE_TTS_DUMP_DECODER
• CRISPASR_VIBEVOICE_TTS_DUMP_PERFRAME
• CRISPASR_VIBEVOICE_TTS_FLASH_ATTN
• CRISPASR_VIBEVOICE_TTS_LATENTS
• CRISPASR_VIBEVOICE_TTS_NOISE
• CRISPASR_VIBEVOICE_TTS_SEED
• CRISPASR_VIBEVOICE_TTS_TRACE
• CRISPASR_VIBEVOICE_TTS_TRACE_FRAME
• CRISPASR_VIBEVOICE_VAE_BACKEND
• CRISPASR_VIBEVOICE_VOICE_AUDIO

VoxCPM2

• CRISPASR_VOXCPM2_BENCH
• CRISPASR_VOXCPM2_BUCKET_CUDA
• CRISPASR_VOXCPM2_CFG_INTERVAL
• CRISPASR_VOXCPM2_CFG_INTERVAL_DEBUG
• CRISPASR_VOXCPM2_CFG_VALUE
• CRISPASR_VOXCPM2_CPU_ONLY
• CRISPASR_VOXCPM2_FA_CPU
• CRISPASR_VOXCPM2_FORCE_SCALAR
• CRISPASR_VOXCPM2_INFERENCE_STEPS
• CRISPASR_VOXCPM2_MAX_LEN
• CRISPASR_VOXCPM2_NAN_CHECK
• CRISPASR_VOXCPM2_NO_BUCKET
• CRISPASR_VOXCPM2_USE_GRAPH — persistent-graph decode. Default ON; 0 opts out.
• CRISPASR_VOXCPM2_USE_REF
• CRISPASR_VOXCPM2_VAE_ENC_DIFF
• CRISPASR_VOXCPM2_VAE_TRACE
• CRISPASR_VOXCPM2_VAE_MAX_SAMPLES - maximum 16 kHz input samples accepted by one voxcpm2-vae upscaling call (default 960000, or 60 seconds). Split longer audio, or raise this only when enough RAM/VRAM is available.

Voxtral / Voxtral-TTS

• CRISPASR_VOXTRAL_BENCH
• CRISPASR_VOXTRAL_FUSED_QKV
• CRISPASR_VOXTRAL_TTS_CODEC_FROM_FILE
• CRISPASR_VOXTRAL_TTS_DEBUG
• CRISPASR_VOXTRAL_TTS_DIFF_DUMP
• CRISPASR_VOXTRAL_TTS_FM_STEPS
• CRISPASR_VOXTRAL_TTS_SEMANTIC_CB
• CRISPASR_VOXTRAL_TTS_TEXT
• CRISPASR_VOXTRAL_TTS_TIMING
• CRISPASR_VOXTRAL_TTS_VOICE

Voxtral-4B

• CRISPASR_VOXTRAL4B_BENCH
• CRISPASR_VOXTRAL4B_FUSED_QKV
• CRISPASR_VOXTRAL4B_STREAM_BATCH_ENCODER
• CRISPASR_VOXTRAL4B_STREAM_CHUNK_MS
• CRISPASR_VOXTRAL4B_STREAM_DEBUG
• CRISPASR_VOXTRAL4B_STREAM_DECODER_THREAD
• CRISPASR_VOXTRAL4B_STREAM_DIFF
• CRISPASR_VOXTRAL4B_STREAM_LIVE
• CRISPASR_VOXTRAL4B_STREAM_TIMING

VAD (encoder/decoder Silero-style)

• CRISPASR_VAD_ENCDEC_CONV_CAST
• CRISPASR_VAD_ENCDEC_CPU
• CRISPASR_VAD_ENCDEC_DEBUG
• CRISPASR_VAD_ENCDEC_PERSIST
• CRISPASR_VAD_ENCDEC_SERIAL_MEL

VAD (WebRTC)

• CRISPASR_WEBRTC_VAD_MODE — WebRTC VAD aggressiveness, 0–3 (default 1). Only consulted when the caller did not pass an explicit mode.

Wav2Vec2

• CRISPASR_WAV2VEC2_BENCH
• CRISPASR_WAV2VEC2_DUMP_DIR
• CRISPASR_WAV2VEC2_VERBOSE

Whisper (Tiron speaker attribution)

• CRISPASR_WHISPER_TIRON — 0 forces the stock whisper decode for A/B. The Tiron constrained-decoding grammar is auto-on whenever the model’s vocab has speaker tokens; plain greedy loses ~5 cpWER.
• CRISPASR_WHISPER_TIRON_DEBUG
• CRISPASR_WHISPER_TIRON_MAX_SPEAKERS
• CRISPASR_WHISPER_TIRON_NOSPEECH — 0 disallows an initial <|nospeech|>.

WavTokenizer

• CRISPASR_WAVTOK_BENCH
• CRISPASR_WAVTOK_DUMP_DIR
• CRISPASR_WAVTOK_FIXED_CODES

Zonos

• CRISPASR_ZONOS_CPP_DUMP_DIR
• CRISPASR_ZONOS_DECODE_CHUNK
• CRISPASR_ZONOS_DECODE_CTX
• CRISPASR_ZONOS_DIFF_N_STEPS
• CRISPASR_ZONOS_FASTCONV
• CRISPASR_ZONOS_G2P — builtin | espeak | auto (default). Which phonemizer zonos tries first (#435). espeak is the pre-#435-follow-on cascade bit for bit (in-process libespeak-ng → the espeak-ng binary → raw ASCII → refuse) and never consults the built-in G2P; builtin puts the built-in EN/DE/FR/ES/RU G2P from crispasr-core first; auto keeps espeak first and uses the built-in only as a fallback below it. The default is not builtin: the built-ins emit espeak-dialect IPA while zonos’s inventory comes from its own conditioning.py symbol list, and unmapped codepoints are dropped silently — the failure mode that made #435 look like a working backend. A language’s default moves only on per-language agreement and drop measurements, never on “it produced audio”.
• CRISPASR_ZONOS_RU_DIALECT — native (default) | espeak. Which SPELLING of the Russian phonemes to hand the model when the built-in G2P produces them. native is what crispasr-core’s Russian G2P emits — a narrow transcription (ɐ/ə reduction gradation, ʂ/ʐ retroflexes, lʲ vs ɫ, æ fronting). espeak rewrites the same sounds into espeak-ng’s ru conventions (ʌ, ʃ, ʒ, ɭ, ɑ, y). Not cosmetic and invisible to the drop counter: every symbol on both sides is inside zonos’s inventory, so nothing is dropped either way — but a model conditions on the spelling it was TRAINED on, and zonos was phonemised with espeak. Measured over 2,200 dictionary words phonemised both ways, raw symbol agreement between the two spellings is 57.7% and the conversion takes it to 88.0%. Same class of problem as CRISPASR_KOKORO_MISAKI_IPA (#316), one language further on. Off by default because it was measured and it LOSES. It did raise phoneme-ID agreement with the espeak arm on all three test sentences (0.773/0.759/0.627 → 0.818/0.852/0.847) — and took the ASR roundtrip from 0.293 to 0.000 on every sentence, in both the with-espeak and the espeak-removed runs. Six arms, six zeros; it was the only arm that never produced a recognisable transcript. Agreement with the tool a model was trained on turns out not to be a proxy for the quality of its audio. Kept as a lever because a different consumer (a piper or kokoro Russian voice trained on espeak) may want it, but it must not be enabled for zonos without a new measurement.
• CRISPASR_ZONOS_G2P_DEBUG — 1 prints the phonemisation readout to stderr: which path ran, the IPA, the full phoneme-ID sequence, and how many emitted codepoints zonos’s inventory could not map (with a histogram of which). Off by default; it exists so a comparison between two G2P paths can come back negative instead of “both produced a wav”.
• CRISPASR_ZONOS_SPEAKER_EMB_PATH
• CRISPASR_ZONOS_TTS_BENCH
• CRISPASR_ZONOS_TTS_TEXT
• CRISPASR_ZONOS_VULKAN_NATIVE

CRISPASR_COSYVOICE3_CAMPP_TAIL

legacy makes cosyvoice3’s CAM++ speaker encoder use the old partial-tail divisor. Default is the same convention as every other CAM++ backend.

Both paths synthesise correctly — the TTS→ASR roundtrip is 8/8 on each — so this is a CONSISTENCY switch, not a correctness one. The eight speaker embeddings baked into the shipped cosyvoice3-voices.gguf were produced with the old divisor, so by default a voice cloned from a WAV and the same voice taken from the bank differ by cos ~0.998. Set this to legacy if you need those two paths to agree.

Why cosyvoice3 specifically is unsettled: its upstream is campplus.onnx, and two onnxruntime builds disagree about AveragePool(ceil_mode=1) on the same clip. CRISPASR_CAMPP_LEGACY_SEGPOOL also exists but is GLOBAL — it would drag chatterbox, confucius4, dots-tts and fireredtts3 away from their own settled PyTorch references to answer a cosyvoice3-only question.

This site is open source. Improve this page.