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"))

models = {
    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",
        "whisper-base",
    ]
}


def recognize(audio: tuple[int, np.ndarray]):
    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)

        results = []
        for name, model in models.items():
            start = timer()
            result = model.recognize(waveform, sample_rate=sample_rate, language="ru")
            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


with gr.Blocks() as demo:
    gr.Markdown("""
    # ASR demo using onnx-asr (Russian models)
    **[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`).
    """)
    input = gr.Audio(min_length=1, max_length=20)
    with gr.Row():
        gr.ClearButton(input)
        btn = gr.Button("Recognize", variant="primary")
    output = gr.Dataframe(headers=["model", "result", "time"], wrap=True)
    btn.click(fn=recognize, inputs=input, outputs=output)
    with gr.Accordion("ASR models used in this demo", open=False):
        gr.Markdown("""
        * `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))
        * `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))
        * `whisper-base` - OpenAI Whisper Base exported with onnxruntime ([origin](https://huggingface.co/openai/whisper-base), [onnx](https://huggingface.co/istupakov/whisper-base-onnx))
        """)

demo.launch()