app.py

import os
from huggingface_hub import snapshot_download

MODEL_CACHE_DIR = "./models"
SENSE_VOICE_LOCAL_PATH = os.path.join(MODEL_CACHE_DIR, "SenseVoiceSmall")
PARAFORMER_LOCAL_PATH = os.path.join(MODEL_CACHE_DIR, "paraformer-zh")
VAD_LOCAL_PATH = os.path.join(MODEL_CACHE_DIR, "fsmn-vad")
PUNC_LOCAL_PATH = os.path.join(MODEL_CACHE_DIR, "ct-punc")

os.makedirs(MODEL_CACHE_DIR, exist_ok=True)


def download_if_missing(repo_id, local_path, name):
    if not os.path.exists(local_path):
        print(f"Downloading {name}...")
        snapshot_download(repo_id=repo_id, local_dir=local_path, ignore_patterns=["*.onnx"])
        print(f"{name} ready.")
    else:
        print(f"{name} found locally.")


download_if_missing("FunAudioLLM/SenseVoiceSmall", SENSE_VOICE_LOCAL_PATH, "SenseVoice")
download_if_missing("funasr/paraformer-zh", PARAFORMER_LOCAL_PATH, "Paraformer-zh")
download_if_missing("funasr/fsmn-vad", VAD_LOCAL_PATH, "FSMN-VAD")
download_if_missing("funasr/ct-punc", PUNC_LOCAL_PATH, "CT-Punc")

import gradio as gr
import time
import tempfile
from funasr import AutoModel
from funasr.utils.postprocess_utils import rich_transcription_postprocess

loaded_models = {}


def get_model(pipeline):
    if pipeline in loaded_models:
        return loaded_models[pipeline]

    if pipeline == "sensevoice":
        model = AutoModel(
            model=SENSE_VOICE_LOCAL_PATH,
            vad_model=VAD_LOCAL_PATH,
            vad_kwargs={"max_single_segment_time": 30000},
            device="cpu",
            disable_update=True,
            hub="hf",
        )
    elif pipeline == "paraformer":
        model = AutoModel(
            model=PARAFORMER_LOCAL_PATH,
            vad_model=VAD_LOCAL_PATH,
            punc_model=PUNC_LOCAL_PATH,
            device="cpu",
            disable_update=True,
            hub="hf",
        )
    else:
        raise ValueError(f"Unknown pipeline: {pipeline}")

    loaded_models[pipeline] = model
    return model


def transcribe(audio_input, pipeline_type):
    if audio_input is None:
        return "Please upload or record audio.", ""

    model = get_model(pipeline_type)

    t0 = time.time()
    if pipeline_type == "sensevoice":
        res = model.generate(
            input=audio_input, cache={}, language="auto",
            use_itn=True, batch_size_s=60, merge_vad=True, merge_length_s=15,
        )
    else:
        res = model.generate(input=audio_input)

    text = rich_transcription_postprocess(res[0]["text"])
    elapsed = time.time() - t0

    metrics = f"Time: {elapsed:.2f}s | Model: {pipeline_type} | Device: CPU"
    return metrics, text


with gr.Blocks(title="FunASR Demo") as demo:
    gr.Markdown("""
# FunASR: Speech Recognition Demo

Industrial-grade ASR toolkit. Upload audio and get transcription instantly.

- **SenseVoice**: Multi-task (ASR + emotion + events), 5 languages, ultra-fast
- **Paraformer**: Non-autoregressive Chinese ASR with punctuation

[GitHub](https://github.com/modelscope/FunASR) | [Docs](https://modelscope.github.io/FunASR/) | [pip install funasr](https://pypi.org/project/funasr/)
    """)

    audio_input = gr.Audio(label="Upload or Record Audio", sources=["upload", "microphone"], type="filepath")

    pipeline_type = gr.Dropdown(
        choices=["sensevoice", "paraformer"],
        label="Model",
        value="sensevoice"
    )

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

    metrics_out = gr.Textbox(label="Metrics", lines=1)
    text_out = gr.Textbox(label="Transcription", lines=8)

    btn.click(transcribe, inputs=[audio_input, pipeline_type], outputs=[metrics_out, text_out])

    gr.Markdown("""
### Install & Use Locally
```python
pip install funasr
from funasr import AutoModel
model = AutoModel(model="funasr/paraformer-zh", hub="hf", vad_model="funasr/fsmn-vad", punc_model="funasr/ct-punc")
result = model.generate(input="audio.wav")
```
    """)

demo.queue().launch()