MicroASR for Python

中文 | English

Offline streaming RNN-T speech recognition for Python using the same plain ONNX model packages as the .NET implementation. Input audio is mono signed 16-bit little-endian PCM at exactly 16 kHz.

Models are not included in the Python package. Applications supply a compatible plain ONNX model directory at runtime.

Install

From PyPI:

python -m pip install microasr

For local development from the repository root:

python -m pip install -e ./sdk/python

NumPy and ONNX Runtime are required. Install the optional audio capture backend to use the system command:

python -m pip install -e "./sdk/python[audio]"

Commands

Recognize a WAV file on Windows, Linux, or macOS:

microasr wav models/en-US sample.wav

Read headerless mono PCM16 from standard input:

ffmpeg -i sample.mp3 -ac 1 -ar 16000 -f s16le - | microasr pcm models/en-US

Capture the default system output when the operating system exposes a loopback or monitor source:

microasr system models/en-US

System capture uses SoundCard. Windows normally exposes WASAPI loopback. Linux requires a PulseAudio/PipeWire monitor source. macOS may require a virtual audio device such as BlackHole.

Library usage

from microasr import RecognitionEventType, RnntRecognitionMode, StreamingRecognizer


def on_result(event):
    if event.type is RecognitionEventType.FINAL:
        print(event.text)


with StreamingRecognizer("models/en-US", RnntRecognitionMode.BALANCED) as recognizer:
    recognizer.add_result_handler(on_result)
    recognizer.write(pcm_chunk)

write() processes one PCM chunk synchronously and returns any events produced by that chunk. complete() flushes an active utterance. Partial and final events also go to registered handlers.

Decoder profiles and parameters

The Accuracy, Balanced, and Performance profiles match the .NET package. All decoder and streaming segmentation parameters are public. Start from a profile and use dataclasses.replace to override only the values required by the application.