app.py

import os
import gradio as gr
import spaces
import dolphin
from dolphin.languages import LANGUAGE_CODES, LANGUAGE_REGION_CODES

MODEL_DIR = os.path.join(os.path.dirname(os.path.abspath(__file__)), "models")
os.makedirs(MODEL_DIR, exist_ok=True)

language_options = [(f"{code}: {name[0]}", code)
                    for code, name in LANGUAGE_CODES.items()]
language_options.sort(key=lambda x: x[0])

MODELS = {
    "base (140M)": "base",
    "small (372M)": "small",
}

language_to_regions = {}
for lang_region, names in LANGUAGE_REGION_CODES.items():
    if "-" in lang_region:
        lang, region = lang_region.split("-", 1)
        if lang not in language_to_regions:
            language_to_regions[lang] = []
        language_to_regions[lang].append((f"{region}: {names[0]}", region))



def update_regions(language):
    if language and language in language_to_regions:
        regions = language_to_regions[language]
        regions.sort(key=lambda x: x[0])
        return gr.Dropdown.update(choices=regions, value=regions[0][1], visible=True)
    return gr.Dropdown.update(choices=[], value=None, visible=False)



@spaces.GPU
def transcribe_audio(audio_file, model_name, language, region, predict_timestamps, padding_speech):
    model_key = MODELS[model_name]
    model = dolphin.load_model(model_key, MODEL_DIR, "cuda")

    waveform = dolphin.load_audio(audio_file)

    kwargs = {
        "predict_time": predict_timestamps,
        "padding_speech": padding_speech
    }

    if language:
        kwargs["lang_sym"] = language
        if region:
            kwargs["region_sym"] = region

    result = model(waveform, **kwargs)

    output_text = result.text
    language_detected = f"{result.language}"
    region_detected = f"{result.region}"

    detected_info = f"Detected language: {result.language}" + \
        (f", region: {result.region}" if result.region else "")
    return output_text, detected_info


with gr.Blocks(title="Dolphin Speech Recognition") as demo:
    gr.Markdown("# Dolphin ASR")
    gr.Markdown("""
    A multilingual, multitask ASR model supporting 40 Eastern languages and 22 Chinese dialects.
    
    This model is from [DataoceanAI/Dolphin](https://github.com/DataoceanAI/Dolphin), for speech recognition in 
    Eastern languages including Chinese, Japanese, Korean, and many more.
    """)

    with gr.Row():
        with gr.Column():
            audio_input = gr.Audio(
                type="filepath", label="Upload or Record Audio")

            with gr.Row():
                model_dropdown = gr.Dropdown(
                    choices=list(MODELS.keys()),
                    value=list(MODELS.keys())[1],
                    label="Model Size"
                )

            with gr.Row():
                language_dropdown = gr.Dropdown(
                    choices=language_options,
                    value=None,
                    label="Language (Optional)",
                    info="If not selected, the model will auto-detect language"
                )
                region_dropdown = gr.Dropdown(
                    choices=[],
                    value=None,
                    label="Region (Optional)",
                    visible=False
                )

            with gr.Row():
                timestamp_checkbox = gr.Checkbox(
                    value=True,
                    label="Include Timestamps"
                )
                padding_checkbox = gr.Checkbox(
                    value=True,
                    label="Pad Speech to 30s"
                )

            transcribe_button = gr.Button("Transcribe", variant="primary")

        with gr.Column():
            output_text = gr.Textbox(label="Transcription", lines=10)
            language_info = gr.Textbox(label="Detected Language", lines=1)

    language_dropdown.change(
        fn=update_regions,
        inputs=[language_dropdown],
        outputs=[region_dropdown]
    )

    transcribe_button.click(
        fn=transcribe_audio,
        inputs=[
            audio_input,
            model_dropdown,
            language_dropdown,
            region_dropdown,
            timestamp_checkbox,
            padding_checkbox
        ],
        outputs=[output_text, language_info]
    )

    gr.Examples(
        inputs=[
            audio_input,
            model_dropdown,
            language_dropdown,
            region_dropdown,
            timestamp_checkbox,
            padding_checkbox
        ],
        outputs=[output_text, language_info],
        fn=transcribe_audio,
        cache_examples=True,
    )

    gr.Markdown("""
    
    - The model supports 40 Eastern languages and 22 Chinese dialects
    - You can let the model auto-detect language or specify language and region
    - Timestamps can be included in the output
    - Speech can be padded to 30 seconds for better processing
    
    
    - Model: [DataoceanAI/Dolphin](https://github.com/DataoceanAI/Dolphin)
    - Paper: [Dolphin: A Multilingual Model for Eastern Languages](https://arxiv.org/abs/2503.20212)
    """)

demo.launch()