ASR API (Python)

The collection of gRPC APIs for Techmo ASR solutions supplied as a Python package.

Setup

Run once after cloning to initialise the submodule:

./setup.sh

Requirements

• Python >=3.8

Installation

Virtual environment

Example:

python3 -m venv .venv
source .venv/bin/activate
pip install --require-virtualenv --upgrade pip
pip install --require-virtualenv .

For basic development use, consider convenient ./install.sh.

Running tests

Proto stubs must be generated before running tests. Use ./install.sh once, then invoke tox:

./install.sh
uvx --with "tox-uv>=1" tox

To run a single Python version:

uvx --with "tox-uv>=1" tox -e py312

Usage

Import

The package provides a precompiled collection of .proto files that can be imported directly or through the alias modules.

Example:

• direct import

>>> from techmo.asr.api.v1p1 import asr_pb2 as api
>>> hasattr(api, "StreamingRecognizeRequest")
True

• import from an alias module

>>> from asr_api import v1p1 as api
>>> hasattr(api, "StreamingRecognizeRequest")
True

Invoke RPC

Invoking RPC simply requires to call a desired method on a stub object dedicated to a specific service.

Example:

import wave

import grpc
from asr_api import v1p1 as api

# This example assumes that the endpoint is an instance
# of techmo.asr.api.v1p1.Asr service listening on the local 30384 port.
grpc_service_address = "127.0.0.1:30384"

# Audio data come from a mono 16-bit linear PCM encoded WAV file.
audio_file = wave.open("audio.wav", "rb")
audio_bytes = audio_file.readframes(audio_file.getnframes())
audio_sampling_rate_hz = audio_file.getframerate()

with grpc.insecure_channel(grpc_service_address) as grpc_channel:
    asr_stub = api.AsrStub(grpc_channel)

    requests = (
        api.StreamingRecognizeRequest(
            config=api.StreamingRecognizeRequestConfig(
                audio_config=api.AudioConfig(
                    encoding=api.AudioConfig.AudioEncoding.LINEAR16,
                    sampling_rate_hz=audio_sampling_rate_hz,
                ),
                speech_recognition_config=api.SpeechRecognitionConfig(
                    enable_speech_recognition=True,
                ),
            ),
        ),
        api.StreamingRecognizeRequest(
            data=api.StreamingRecognizeRequestData(
                audio=api.Audio(bytes=audio_bytes),
            ),
        ),
    )

    for response in asr_stub.StreamingRecognize(iter(requests)):
        print(response)

Note about audio bytes and samples

Even though samples are easier to work with, finally audio must be sent in its pure binary representation, not decoded as samples. To encode samples into bytes, the struct Python module comes in handy.

import struct

# Imagine that audio samples are an array of integers
# (mono signed little-endian 16-bit linear PCM encoded samples in this case).
audio_samples = []
audio_bytes = struct.pack(f"<{len(audio_samples)}h", *audio_samples)