voxrt 0.2.1

Fast, deterministic speech-to-text for Python — a Rust runtime for Parakeet Redux.

voxrt runs moondream/parakeet-redux, the 1.58-bit ternary distillation of NVIDIA's Parakeet-TDT 0.6B (25 European languages, 178 MB), with native ternary kernels written in Rust. No PyTorch, ONNX or ML framework: one small wheel and huggingface_hub.

• Fast — 197× real time on LibriSpeech test-clean on an M4 Max GPU, 105× on an 8-core AMD server CPU
• Apple GPU by default on Apple silicon (Metal); the Neural Engine on request; CPU everywhere else
• Same bits everywhere — the CPU path gives byte-identical transcripts on Apple silicon and x86
• Long audio built in — the model's own VAD head cuts talks and meetings at pauses; segment and word timestamps
• Streaming — push audio as it arrives; the streamed transcript equals the offline one
• Accurate — WER 1.95 on LibriSpeech test-clean (the model card reports 1.96)

Install

pip install voxrt            # WAV input
pip install "voxrt[audio]"   # plus FLAC / OGG / MP3 and numpy arrays

Wheels for Apple silicon and Intel Macs (macOS 10.12+) and Linux x86-64, Python 3.9 and later. The model (≈ 180 MB) downloads into the Hugging Face cache on first use.

Quick start

import voxrt

t = voxrt.Transcriber()                       # or voxrt.Transcriber("/path/to/parakeet-redux")
r = t.transcribe("talk.mp3")
print(r.text)

for s in r.segments:                          # Segment(start, end, text, words=[Word(word, start, end)])
    print(f"[{s.start:6.2f} → {s.end:6.2f}] {s.text}")

Audio: WAV is decoded natively, FLAC / OGG / MP3 with voxrt[audio], anything else (M4A, AAC, video) when ffmpeg is on PATH. Arrays work too: t.transcribe(samples, sample_rate=48000) takes mono or [n, channels].

Streaming

st = t.stream()
for chunk in microphone_chunks_16k():         # float32, 16 kHz mono
    for seg in st.push(chunk):                # segments are emitted as soon as their end is certain
        print(seg.text)
for seg in st.flush():
    print(seg.text)

Devices

voxrt.Transcriber(device="auto")   # default: the Apple GPU on Apple silicon, the CPU elsewhere
voxrt.Transcriber(device="cpu")    # byte-identical across machines
voxrt.Transcriber(device="ane")    # Apple Neural Engine (macOS 15+): downloads a Core ML bundle, ≈ 300 MB

t.device reports where the encoder runs. The Neural Engine bundle, MaksL/voxrt-parakeet-redux-coreml, keeps the ternary weights exact in 2 bits; macOS compiles it on the first load (≈ 5 min, once).

Results are frozen dataclasses (Transcript, Segment, Word) with type hints. The GIL is released while transcribing, so one Transcriber can serve several threads.

How it works

Benchmarks

Apple M4 Max, the same method for every runtime (processing time, model load excluded):

More machines, the method and the comparison with other runtimes: benchmarks in the repository.

Links

• Source, CLI, Swift and C bindings: github.com/lmaksym/voxrt
• Model: moondream/parakeet-redux · Neural Engine bundle: MaksL/voxrt-parakeet-redux-coreml
• Changes: CHANGELOG

Development

Build from source with maturin develop --release in bindings/python. Tests: pip install "voxrt[test]", then VOXRT_TEST_DOWNLOAD=1 pytest bindings/python/tests (or point VOXRT_MODEL_DIR at a local copy). The test clip is LibriSpeech test-clean 1089-134686-0001 (CC BY 4.0, see tests/audio/ATTRIBUTION.md).

License

voxrt is Apache-2.0. The model weights are © NVIDIA and M87 Labs (Moondream), CC-BY-4.0.