app.py

import uuid
import threading

import gradio as gr
import numpy as np
import sherpa_onnx
from huggingface_hub import hf_hub_download


# ============================================================
# CONFIG
# ============================================================

MODEL_REPO = "hynt/Zipformer-30M-RNNT-Streaming-6000h"

SAMPLE_RATE = 16000

STREAM_EVERY = 0.25


# ============================================================
# DOWNLOAD MODEL
# ============================================================

print("Downloading Zipformer model...")

encoder_path = hf_hub_download(
    repo_id=MODEL_REPO,
    filename="encoder-epoch-31-avg-11-chunk-32-left-128.fp16.onnx",
)

decoder_path = hf_hub_download(
    repo_id=MODEL_REPO,
    filename="decoder-epoch-31-avg-11-chunk-32-left-128.fp16.onnx",
)

joiner_path = hf_hub_download(
    repo_id=MODEL_REPO,
    filename="joiner-epoch-31-avg-11-chunk-32-left-128.fp16.onnx",
)

tokens_path = hf_hub_download(
    repo_id=MODEL_REPO,
    filename="config.json",
)

print("Model downloaded.")


# ============================================================
# CREATE ONLINE RECOGNIZER
# ============================================================

print("Creating sherpa-onnx OnlineRecognizer...")

recognizer = sherpa_onnx.OnlineRecognizer.from_transducer(
    tokens=tokens_path,
    encoder=encoder_path,
    decoder=decoder_path,
    joiner=joiner_path,
    num_threads=2,
    sample_rate=SAMPLE_RATE,
    feature_dim=80,
    decoding_method="greedy_search",
    provider="cpu",
)

print("OnlineRecognizer ready.")


# ============================================================
# SERVER-SIDE SESSION STATE
#
# KHÔNG lưu OnlineStream trong gr.State.
# OnlineStream là C++ object và không pickle được.
#
# gr.State chỉ giữ session_id dạng string.
# ============================================================

STREAMS = {}

STREAM_LOCK = threading.Lock()


def create_session_id():
    return str(uuid.uuid4())


def get_or_create_stream(session_id):
    if not session_id:
        session_id = create_session_id()

    with STREAM_LOCK:
        if session_id not in STREAMS:
            print(f"[SESSION] Create stream: {session_id}")

            STREAMS[session_id] = {
                "stream": recognizer.create_stream(),
                "callbacks": 0,
            }

        state = STREAMS[session_id]

    return session_id, state


def remove_stream(session_id):
    if not session_id:
        return

    with STREAM_LOCK:
        if session_id in STREAMS:
            print(f"[SESSION] Remove stream: {session_id}")
            del STREAMS[session_id]


# ============================================================
# AUDIO PREPROCESSING
# ============================================================

def preprocess_audio(audio):
    """
    Gradio type='numpy' returns:

        (
            sample_rate,
            np.ndarray
        )

    Return:
        sample_rate,
        mono float32 waveform
    """

    if audio is None:
        return None, None

    sample_rate, samples = audio

    samples = np.asarray(samples)

    # --------------------------------------------------------
    # Stereo -> mono
    # --------------------------------------------------------

    if samples.ndim == 2:
        samples = samples.mean(axis=1)

    # --------------------------------------------------------
    # Integer PCM -> float32 [-1, 1]
    # --------------------------------------------------------

    if np.issubdtype(samples.dtype, np.integer):

        dtype_info = np.iinfo(samples.dtype)

        max_value = max(
            abs(dtype_info.min),
            dtype_info.max,
        )

        samples = (
            samples.astype(np.float32)
            / max_value
        )

    else:
        samples = samples.astype(np.float32)

    # Safety
    samples = np.nan_to_num(
        samples,
        nan=0.0,
        posinf=0.0,
        neginf=0.0,
    )

    return sample_rate, samples


# ============================================================
# STREAMING TRANSCRIPTION
# ============================================================

def transcribe_stream(
    audio,
    session_id,
):
    """
    Called repeatedly by Gradio while microphone is recording.

    Each callback:
        new audio chunk
            ↓
        same OnlineStream
            ↓
        decode all ready frames
            ↓
        partial transcript
    """

    session_id, state = get_or_create_stream(
        session_id
    )

    stream = state["stream"]

    if audio is None:
        return session_id, ""

    sample_rate, samples = preprocess_audio(
        audio
    )

    if samples is None or samples.size == 0:
        return session_id, ""

    state["callbacks"] += 1

    callback_id = state["callbacks"]

    duration = (
        samples.shape[0]
        / sample_rate
    )

    print(
        f"[AUDIO #{callback_id}] "
        f"sr={sample_rate} "
        f"samples={samples.shape[0]} "
        f"duration={duration:.3f}s"
    )

    # ========================================================
    # IMPORTANT:
    #
    # Feed the CURRENT microphone chunk to the SAME
    # sherpa OnlineStream.
    #
    # sherpa-onnx can resample e.g. 48 kHz -> 16 kHz.
    # ========================================================

    stream.accept_waveform(
        sample_rate,
        samples,
    )

    # ========================================================
    # DECODE AVAILABLE FRAMES
    # ========================================================

    decode_count = 0

    while recognizer.is_ready(stream):

        recognizer.decode_stream(
            stream
        )

        decode_count += 1

    # ========================================================
    # PARTIAL HYPOTHESIS
    # ========================================================

    result = recognizer.get_result(
        stream
    )

    result = (
        result.strip()
        if result
        else ""
    )

    print(
        f"[ASR #{callback_id}] "
        f"decode_calls={decode_count} "
        f"text={result!r}"
    )

    return (
        session_id,
        result,
    )


# ============================================================
# RESET
# ============================================================

def reset_stream(session_id):

    remove_stream(
        session_id
    )

    new_session_id = (
        create_session_id()
    )

    print(
        f"[RESET] New session: "
        f"{new_session_id}"
    )

    return (
        new_session_id,
        "",
    )


# ============================================================
# GRADIO UI
# ============================================================

with gr.Blocks(
    title="Vietnamese Zipformer Streaming ASR"
) as demo:

    gr.Markdown(
        """
# Vietnamese Zipformer Streaming ASR

**Model:** `hynt/Zipformer-30M-RNNT-Streaming-6000h`

**Architecture:** Zipformer + RNN-Transducer

**Runtime:** sherpa-onnx

Bấm microphone và nói tiếng Việt. Transcript sẽ được cập nhật
liên tục khi model có partial hypothesis.
"""
    )

    # --------------------------------------------------------
    # Only store a serializable string in gr.State
    # --------------------------------------------------------

    session_id = gr.State(
        value=None
    )

    # --------------------------------------------------------
    # IMPORTANT: streaming=True
    # --------------------------------------------------------

    microphone = gr.Audio(
        sources=["microphone"],
        type="numpy",
        streaming=True,
        label="Microphone",
    )

    transcript = gr.Textbox(
        label="Live Transcript",
        lines=6,
        interactive=False,
    )

    reset_button = gr.Button(
        "Reset"
    )

    # --------------------------------------------------------
    # Streaming event
    # --------------------------------------------------------

    microphone.stream(
        fn=transcribe_stream,

        inputs=[
            microphone,
            session_id,
        ],

        outputs=[
            session_id,
            transcript,
        ],

        stream_every=STREAM_EVERY,

        time_limit=120,

        concurrency_limit=1,
    )

    # --------------------------------------------------------
    # Reset session
    # --------------------------------------------------------

    reset_button.click(
        fn=reset_stream,

        inputs=[
            session_id,
        ],

        outputs=[
            session_id,
            transcript,
        ],
    )


# ============================================================
# START APP
# ============================================================

if __name__ == "__main__":
    demo.launch(
        ssr_mode=False
    )