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")

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).

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)

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 run/stream at a time per Model in 0.x: sessions share the model's compute backend, so serialize runs across sessions (or load one model per worker). See the Model docstring.
• Import package: transcribe_cpp
• Distribution: transcribe-cpp
• License: MIT