transcribe-cpp

Python bindings for transcribe.cpp, a C/C++ speech-to-text library built on ggml.

Status: in development. Until wheels are published, use a locally built libtranscribe through repo auto-discovery or TRANSCRIBE_LIBRARY.

Upgrading from 0.1? See the 0.2 migration guide, including the replacement of gpu_device= with exact device objects.

import transcribe_cpp

with transcribe_cpp.Model("model.gguf") as model:
    with model.session() as session:
        result = session.run(pcm_float32_16k_mono)
        print(result.text)

run() takes mono 16 kHz float32 PCM (buffer-protocol object or sequence). It does not decode containers or resample; convert audio before calling it.

import numpy as np

pcm = np.asarray(audio, dtype=np.float32)   # 1-D, 16 kHz mono
# Downmix stereo first; 2-D input is rejected:
# pcm = audio.mean(axis=1).astype(np.float32)
result = session.run(pcm)

Punctuation, capitalization, and text normalization

Generic run controls use "default" to preserve each model family's shipped behavior. Models advertising model.supports("pnc") accept pnc="off" or pnc="on"; models advertising model.supports("itn") accept the equivalent itn values. The options are available on run(), run_batch(), stream(), and the one-shot transcribe() helper.

result = session.run(pcm, pnc="off", itn="on")

Prompting

vocabulary (custom terms), prompt (context, or the instruction under task="instruct") and prefix (text the model continues from) take effect where model.supports() reports "vocabulary", "context_prompt", "instruct" or "transcript_prefix".

result = session.run(pcm, vocabulary=["Kubernetes", "gRPC"])

Streaming models expose incremental transcription with committed/tentative text views — see examples/stream_wav.py:

with model.session() as session, session.stream() as stream:
    for chunk in pcm_chunks:
        stream.feed(chunk)
        text = stream.text()        # .committed (stable) + .tentative
    stream.finalize()
    result = stream.snapshot()      # language, segments, words, tokens, timings

Long transcriptions can be cancelled from another thread with session.cancel() — the run raises Aborted with the partial transcript on exc.partial_result (same for OutputTruncated).

Diarization

Models whose model.roles include Role.DIARIZE (Sortformer) answer "who spoke when" through a diarize session. run() returns SpeakerSegment rows (speaker_id in 1..model.diarize_info.max_speakers), grouped by speaker. Locking, Busy, cancel() and close() work as on Session. A model without the role raises UnsupportedRole, as model.session() and model.capabilities do on a model without Role.ASR (Sortformer serves only Role.DIARIZE).

with model.diarize_session() as diarizer:
    turns = diarizer.run(pcm, family=transcribe_cpp.SortformerDiarizeOptions(
        preset="very_high_latency"))
    for turn in turns:
        print(turn.speaker_id, turn.t0_ms, turn.t1_ms)

Language ID

Models whose model.roles include Role.LANGID (VoxLingua107 ECAPA-TDNN) identify the spoken language through a langid session. run() returns a LangIdResult with candidates ranked by p. Codes are the model's own labels ("iw", "jw"; model.langid_label_index("he") resolves aliases), so match result.code against the ASR model's capabilities.languages yourself. allowed=None scores every label; allowed_mass is the unrestricted probability the allowed set captured. An empty allowed list raises InvalidArgument; clips under model.langid_info.min_audio_ms raise InputTooShort, and longer ones than max_audio_ms are scored on their first max_audio_ms. Every allowed label comes back as a candidate. Locking, Busy, cancel() and close() work as on Session.

with model.langid_session() as lid:
    result = lid.run(pcm, allowed=["en", "de", "fr"])
    print(result.code, result.candidates[0].p, result.allowed_mass)

Voice activity detection

Models whose model.roles include Role.VAD (Silero VAD, GGUF or the whisper.cpp .bin) score speech probability per frame of model.vad_info.frame_samples samples through a VAD session. run() scores one clip and returns a VadResult with probs and speech segments (start_sample/end_sample; options default to Silero's get_speech_timestamps). stream_feed() streams audio of any chunk size and returns the probabilities of the frames each call completed (first_frame is the position of the first); stream_flush() scores the zero-padded remainder and ends the stream, stream_reset() drops it. The VAD stream does not take the model's ASR stream lease. Locking, Busy, cancel() and close() work as on Session.

with model.vad_session() as vad:
    for seg in vad.run(pcm, min_silence_ms=300).segments:
        print(seg.start_sample, seg.end_sample)

    frame = model.vad_info.frame_samples
    probs = []
    for chunk in chunks:
        r = vad.stream_feed(chunk)
        print(f"{len(r.probs)} frames from sample {r.first_frame * frame}")
        probs.extend(r.probs)
    probs.extend(vad.stream_flush().probs)

Backends

Model(backend=...) applies a backend policy ("auto" uses the best available). transcribe_cpp.backends() returns process-local device objects; pass one as Model(device=device) for exact selection with no fallback. Persist a device's device_id, not its runtime handle or index. backend_available(kind) checks whether a backend policy can currently be satisfied.

device = next(d for d in transcribe_cpp.backends() if d.device_type == "cpu")
with transcribe_cpp.Model("model.gguf", device=device) as model:
    print(model.device)

Weights are memory-mapped from the model file where the device can use host memory (CPU, Metal on Apple silicon), so they live in reclaimable page cache rather than process memory. Don't truncate or rewrite a model file in place while it is loaded; write a new file and rename it over the old one. Model(no_mmap=True) copies the weights instead.

Planned wheels will bundle CPU plus platform accelerators; transcribe-cpp[cu12] will add the CUDA 12 provider.

Running from a working tree

The binding loads the native library at import and verifies its ABI layout and version before use. Build a shared library, then run from the repo or point TRANSCRIBE_LIBRARY at it:

cmake -B build-shared -DTRANSCRIBE_BUILD_SHARED=ON
cmake --build build-shared --target transcribe

cd bindings/python
PYTHONPATH=src uv run --no-project python examples/transcribe_wav.py \
    ../../models/whisper-tiny.en/whisper-tiny.en-Q5_K_M.gguf ../../samples/jfk.wav

No-model tests always run; model tests skip unless smoke assets are present. Override paths with TRANSCRIBE_SMOKE_MODEL, TRANSCRIBE_SMOKE_AUDIO, and TRANSCRIBE_SMOKE_STREAMING_MODEL.

cd bindings/python
TRANSCRIBE_LIBRARY=../../build-shared/src/libtranscribe.dylib \
    uv run --extra test pytest

Notes

• One compute call at a time per Model: the binding serializes calls across all sessions of a model with a model-wide lock (load one model per worker for parallelism). While a stream is active, other runs and stream begins on that model raise transcribe_cpp.Busy. See the Model docstring.
• Import package: transcribe_cpp
• Distribution: transcribe-cpp
• License: MIT