crispasr

Python bindings for CrispASR — lightweight on-device speech recognition via ggml.

Supports the ASR backends compiled into the linked CrispASR library, including Whisper, Qwen3-ASR, FastConformer, Canary, Parakeet, Cohere, Granite-Speech, Voxtral, wav2vec2, GLM-ASR, Kyutai-STT, Moonshine, FireRed, OmniASR, VibeVoice-ASR, and Index-Echo.

The package also exposes TTS, speech-to-speech, music analysis, ContentVec voice conversion, alignment, diarization, VAD, text processors and model utilities. See the Python API inventory for entry points, result units, ownership, callback rules and verification limits. These additions require a matching native build; current source coverage does not imply an older published wheel contains the same APIs.

Install

pip install crispasr

Platform wheels bundle the native libcrispasr — nothing else to install — for Linux (x86_64, arm64), macOS (Apple Silicon, Metal-accelerated), and Windows (x86_64).

GPU wheels

CUDA and Vulkan builds are published to a separate index (llama-cpp-python style — pass --extra-index-url):

# NVIDIA CUDA (Linux + Windows)
pip install crispasr --extra-index-url https://crispstrobe.github.io/CrispASR/whl/cuda/
# Vulkan (Windows)
pip install crispasr --extra-index-url https://crispstrobe.github.io/CrispASR/whl/vulkan/

Other platforms / bring-your-own library

Where no prebuilt wheel matches, pip installs the pure-Python sdist, which loads a libcrispasr you supply. Build/install it from source and, if it lands in a non-standard location, point CRISPASR_LIB_PATH at the file:

git clone https://github.com/CrispStrobe/CrispASR
cd CrispASR && cmake -B build && cmake --build build -j && sudo cmake --install build
export CRISPASR_LIB_PATH=/usr/local/lib/libcrispasr.so   # only if non-standard

Quick start

from crispasr import CrispASR

model = CrispASR("ggml-base.en.bin")
for seg in model.transcribe("audio.wav"):
    print(f"[{seg.start:.1f}s - {seg.end:.1f}s] {seg.text}")
model.close()

Or use the unified Session API for non-Whisper backends (Qwen3-ASR, FastConformer, Parakeet, …):

from crispasr import Session

s = Session("qwen3-asr-0.6b-q4_k.gguf")
for seg in s.transcribe(pcm_f32, sample_rate=16000):
    print(seg.text)

Mono array inputs at other sample rates are converted to 16 kHz with a band-limited Kaiser polyphase filter. Older libraries use the matching NumPy fallback, without a SciPy dependency. Strided arrays are made contiguous before native inference. See audio resampling for limits and regression checks.

Index-Echo uses the same Session API. Open the tower GGUF with its matching decoder beside it, then call s.set_target_language("en") (also ja or es). Each segment contains transcript\ntranslation and generated subtitle times. Place Silero v6.2 beside the pair for the released full-file window recipe.

Session segments expose tokens (SessionToken: text, start/end seconds, confidence, and lossless raw_text bytes). Missing word or token timestamps are None; check word.has_timestamps before doing arithmetic. VibeVoice has native segment times but no native word times. Use align_words() for word alignment, then free_aligner_cache() or session.free_aligner_cache() to release the process-wide aligner allocation. Older libraries return empty token lists and report cache release as unavailable. See session metadata and compatibility. Parakeet metadata is retained through Japanese slice/gap-fill and legacy overlap selection; raw token blocks follow retained words even when timestamps are tied. These main-branch changes require the matching native library and are not in v0.8.43.

API

• CrispASR — Whisper-compatible high-level API
• Session — unified API across all backends compiled into libcrispasr
• ChatSession — text → text chat over a GGUF chat model: one-shot and streaming generation, prompt-token counting, and cancellation through an abort predicate
• align_words(...) — word-level CTC alignment
• free_aligner_cache(...) — release the process-wide forced-aligner cache
• diarize_segments(...) — speaker diarization (energy / xcorr / vad-turns / pyannote / FoxNose)
• diarize_segments_with_turns(...) — label segments and return FoxNose's finer audio-derived speaker turns
• SpeakerEmbedder(spec) — pluggable embedder ("auto"/"titanet", "indextts"/"ecapa", or a .gguf path)
• PyannoteCache(pcm, model) — pre-computed pyannote-seg posteriors for cross-slice consistency
• agglomerative_cluster(embeddings, ...) — single-linkage cosine clustering for globally stable speaker IDs
• TitaNet / SpeakerDB — standalone speaker verification + closed-roster profile matching (consent-gated; requires expected_names + consent=True, see docs/diarization-speakers.md)
• detect_language_pcm(...) — language ID
• registry_lookup(...) — auto-download known models from the model hub
• registry_default_bundle(...) — enumerate the exact primary, companion, and extra files used by -m auto, including licence-acceptance policy

See the main repo for full documentation, model registry, and CLI.

License

MIT — see LICENSE.