app.py

import logging
from importlib.metadata import version
from timeit import default_timer as timer

import gradio as gr
import numpy as np

import onnx_asr

logging.basicConfig(format="%(asctime)s %(levelname)s %(message)s", level=logging.WARNING)
logger = logging.getLogger(__name__)
logger.setLevel(logging.DEBUG)
logger.info("onnx_asr version: %s", version("onnx_asr"))

vad = onnx_asr.load_vad("silero")

whisper = {name: onnx_asr.load_model(name) for name in ["whisper-base"]}

models_ru = {
    name: onnx_asr.load_model(name)
    for name in [
        "gigaam-v2-ctc",
        "gigaam-v2-rnnt",
        "nemo-fastconformer-ru-ctc",
        "nemo-fastconformer-ru-rnnt",
        "alphacep/vosk-model-ru",
        "alphacep/vosk-model-small-ru",
    ]
}

models_en = {
    name: onnx_asr.load_model(name)
    for name in [
        "nemo-parakeet-ctc-0.6b",
        "nemo-parakeet-rnnt-0.6b",
        "nemo-parakeet-tdt-0.6b-v2",
    ]
}

models_vad = whisper | models_ru | models_en


def recognize(audio: tuple[int, np.ndarray], models, language):
    if audio is None:
        return None

    sample_rate, waveform = audio
    length = waveform.shape[-1] / sample_rate
    logger.debug("recognize: length %.3f, sample_rate %s, waveform.shape %s.", length, sample_rate, waveform.shape)
    try:
        waveform = waveform.astype(np.float32) / 2 ** (8 * waveform.itemsize - 1)
        if waveform.ndim == 2:
            waveform = waveform.mean(axis=1)

        results = []
        for name, model in models.items():
            if length > 20 and name == "alphacep/vosk-model-small-ru":
                continue
            start = timer()
            result = model.recognize(waveform, sample_rate=sample_rate, language=language)
            time = timer() - start
            logger.debug("recognized by %s: result '%s', time %.3f s.", name, result, time)
            results.append([name, result, f"{time:.3f} s."])

    except Exception as e:
        raise gr.Error(f"{e} Audio: sample_rate: {sample_rate}, waveform.shape: {waveform.shape}.") from e
    else:
        return results


def recognize_ru(audio: tuple[int, np.ndarray]):
    return recognize(audio, models_ru | whisper, "ru")


def recognize_en(audio: tuple[int, np.ndarray]):
    return recognize(audio, models_en | whisper, "en")


def recognize_with_vad(audio: tuple[int, np.ndarray], name: str):
    if audio is None:
        return None

    sample_rate, waveform = audio
    logger.debug("recognize: sample_rate %s, waveform.shape %s.", sample_rate, waveform.shape)
    try:
        waveform = waveform.astype(np.float32) / 2 ** (8 * waveform.itemsize - 1)
        if waveform.ndim == 2:
            waveform = waveform.mean(axis=1)

        model = models_vad[name].with_vad(vad, batch_size=1)
        results = ""
        for res in model.recognize(waveform, sample_rate=sample_rate):
            logger.debug("recognized by %s: result '%s'.", name, res)
            results += f"[{res.start:5.1f}, {res.end:5.1f}]: {res.text}\n"
            yield results

    except Exception as e:
        raise gr.Error(f"{e} Audio: sample_rate: {sample_rate}, waveform.shape: {waveform.shape}.") from e


with gr.Blocks() as recognize_short:
    audio = gr.Audio(min_length=1, max_length=30)
    with gr.Row():
        gr.ClearButton(audio)
        btn_ru = gr.Button("Recognize (ru)", variant="primary")
        btn_en = gr.Button("Recognize (en)", variant="primary")
    output = gr.Dataframe(headers=["model", "result", "time"], wrap=True)
    btn_ru.click(fn=recognize_ru, inputs=audio, outputs=output)
    btn_en.click(fn=recognize_en, inputs=audio, outputs=output)


with gr.Blocks() as recognize_long:
    name = gr.Dropdown(models_vad.keys(), label="Model")
    # lang = gr.Label()
    audio = gr.Audio(min_length=1, max_length=300)
    with gr.Row():
        gr.ClearButton(audio)
        btn = gr.Button("Recognize", variant="primary")
    output = gr.TextArea(label="result")
    btn.click(fn=recognize_with_vad, inputs=[audio, name], outputs=output)

    def on_model_change(name: str):
        if name in models_ru:
            label = f"Model {name} support only Russian language"
        elif name in models_en:
            label = f"Model {name} support only English language"
        else:
            label = None
        return gr.Audio(min_length=1, max_length=300, label=label)

    name.change(on_model_change, inputs=name, outputs=audio)

with gr.Blocks() as demo:
    gr.Markdown("""
    # ASR demo using onnx-asr
    **[onnx-asr](https://github.com/istupakov/onnx-asr)** is a Python package for Automatic Speech Recognition using ONNX models.
    The package is written in pure Python with minimal dependencies (no `pytorch` or `transformers`).
    """)
    gr.TabbedInterface(
        [recognize_short, recognize_long],
        [
            "Recognition of a short phrase (up to 30 sec.)",
            "Recognition of a long phrase with VAD (up to 5 min.)",
        ],
    )
    with gr.Accordion("Models used in this demo:", open=False):
        gr.Markdown("""
        ## Russian ASR models
        * `gigaam-v2-ctc` - Sber GigaAM v2 CTC ([origin](https://github.com/salute-developers/GigaAM), [onnx](https://huggingface.co/istupakov/gigaam-v2-onnx))
        * `gigaam-v2-rnnt` - Sber GigaAM v2 RNN-T ([origin](https://github.com/salute-developers/GigaAM), [onnx](https://huggingface.co/istupakov/gigaam-v2-onnx))
        * `nemo-fastconformer-ru-ctc` - Nvidia FastConformer-Hybrid Large (ru) with CTC decoder ([origin](https://huggingface.co/nvidia/stt_ru_fastconformer_hybrid_large_pc), [onnx](https://huggingface.co/istupakov/stt_ru_fastconformer_hybrid_large_pc_onnx))
        * `nemo-fastconformer-ru-rnnt` - Nvidia FastConformer-Hybrid Large (ru) with RNN-T decoder ([origin](https://huggingface.co/nvidia/stt_ru_fastconformer_hybrid_large_pc), [onnx](https://huggingface.co/istupakov/stt_ru_fastconformer_hybrid_large_pc_onnx))
        * `whisper-base` - OpenAI Whisper Base exported with onnxruntime ([origin](https://huggingface.co/openai/whisper-base), [onnx](https://huggingface.co/istupakov/whisper-base-onnx))
        * `alphacep/vosk-model-ru` - Alpha Cephei Vosk 0.54-ru ([origin](https://huggingface.co/alphacep/vosk-model-ru))
        * `alphacep/vosk-model-small-ru` - Alpha Cephei Vosk 0.52-small-ru ([origin](https://huggingface.co/alphacep/vosk-model-small-ru))
        ## English ASR models
        * `nemo-parakeet-ctc-0.6b` - Nvidia Parakeet CTC 0.6B (en) ([origin](https://huggingface.co/nvidia/parakeet-ctc-0.6b), [onnx](https://huggingface.co/istupakov/parakeet-ctc-0.6b-onnx))
        * `nemo-parakeet-rnnt-0.6b` - Nvidia Parakeet RNNT 0.6B (en) ([origin](https://huggingface.co/nvidia/parakeet-rnnt-0.6b), [onnx](https://huggingface.co/istupakov/parakeet-rnnt-0.6b-onnx))
        * `nemo-parakeet-tdt-0.6b-v2` - Nvidia Parakeet TDT 0.6B V2 (en) ([origin](https://huggingface.co/nvidia/parakeet-tdt-0.6b-v2), [onnx](https://huggingface.co/istupakov/parakeet-tdt-0.6b-v2-onnx))
        * `whisper-base` - OpenAI Whisper Base exported with onnxruntime ([origin](https://huggingface.co/openai/whisper-base), [onnx](https://huggingface.co/istupakov/whisper-base-onnx))
        ## VAD models
        * `silero` - Silero VAD ([origin](https://github.com/snakers4/silero-vad), [onnx](https://huggingface.co/onnx-community/silero-vad))
        """)

demo.launch()