app.py

from transformers import AutoModelForSpeechSeq2Seq, AutoProcessor, pipeline
from transformers.utils import is_flash_attn_2_available
from transformers.pipelines.audio_utils import ffmpeg_read
import torch
import gradio as gr
import time

BATCH_SIZE = 16
MAX_AUDIO_MINS = 30  # maximum audio input in minutes

device = "cuda:0" if torch.cuda.is_available() else "cpu"
torch_dtype = torch.float16 if torch.cuda.is_available() else torch.float32
use_flash_attention_2 = is_flash_attn_2_available()

model = AutoModelForSpeechSeq2Seq.from_pretrained(
    "openai/whisper-large-v2", torch_dtype=torch_dtype, low_cpu_mem_usage=True, use_safetensors=True, use_flash_attention_2=use_flash_attention_2
)
distilled_model = AutoModelForSpeechSeq2Seq.from_pretrained(
    "distil-whisper/distil-large-v2", torch_dtype=torch_dtype, low_cpu_mem_usage=True, use_safetensors=True, use_flash_attention_2=use_flash_attention_2
)

if not use_flash_attention_2:
    # use flash attention from pytorch sdpa
    model = model.to_bettertransformer()
    distilled_model = distilled_model.to_bettertransformer()

processor = AutoProcessor.from_pretrained("openai/whisper-large-v2")

model.to(device)
distilled_model.to(device)

pipe = pipeline(
    "automatic-speech-recognition",
    model=model,
    tokenizer=processor.tokenizer,
    feature_extractor=processor.feature_extractor,
    max_new_tokens=128,
    chunk_length_s=30,
    torch_dtype=torch_dtype,
    device=device,
    generate_kwargs={"language": "en", "task": "transcribe"},
)
pipe_forward = pipe._forward

distil_pipe = pipeline(
    "automatic-speech-recognition",
    model=distilled_model,
    tokenizer=processor.tokenizer,
    feature_extractor=processor.feature_extractor,
    max_new_tokens=128,
    chunk_length_s=15,
    torch_dtype=torch_dtype,
    device=device,
    generate_kwargs={"language": "en", "task": "transcribe"},
)
distil_pipe_forward = distil_pipe._forward

def transcribe(inputs):
    if inputs is None:
        raise gr.Error("No audio file submitted! Please record or upload an audio file before submitting your request.")

    with open(inputs, "rb") as f:
        inputs = f.read()

    inputs = ffmpeg_read(inputs, pipe.feature_extractor.sampling_rate)
    audio_length_mins = len(inputs) / pipe.feature_extractor.sampling_rate / 60

    if audio_length_mins > MAX_AUDIO_MINS:
        raise gr.Error(
            f"To ensure fair usage of the Space, the maximum audio length permitted is {MAX_AUDIO_MINS} minutes."
            f"Got an audio of length {round(audio_length_mins, 3)} minutes."
        )

    inputs = {"array": inputs, "sampling_rate": pipe.feature_extractor.sampling_rate}

    def _forward_distil_time(*args, **kwargs):
        global distil_runtime
        start_time = time.time()
        result = distil_pipe_forward(*args, **kwargs)
        distil_runtime = time.time() - start_time
        distil_runtime = round(distil_runtime, 2)
        return result

    distil_pipe._forward = _forward_distil_time
    distil_text = distil_pipe(inputs, batch_size=BATCH_SIZE)["text"]
    yield distil_text, distil_runtime, None, None, None

    def _forward_time(*args, **kwargs):
        global runtime
        start_time = time.time()
        result = pipe_forward(*args, **kwargs)
        runtime = time.time() - start_time
        runtime = round(runtime, 2)
        return result

    pipe._forward = _forward_time
    text = pipe(inputs, batch_size=BATCH_SIZE)["text"]

    yield distil_text, distil_runtime, text, runtime

if __name__ == "__main__":
    with gr.Blocks() as demo:
        gr.HTML(
            """
                <div style="text-align: center; max-width: 700px; margin: 0 auto;">
                  <div
                    style="
                      display: inline-flex; align-items: center; gap: 0.8rem; font-size: 1.75rem;
                    "
                  >
                    <h1 style="font-weight: 900; margin-bottom: 7px; line-height: normal;">
                      Whisper vs Distil-Whisper
                    </h1>
                  </div>
                </div>
            """
        )
        gr.HTML(
            f"""
            This demo shows a speed comparison between <a href="https://huggingface.co/openai/whisper-large-v2"> Whisper </a> 
            and <a href="https://huggingface.co/distil-whisper/distil-large-v2"> Distil-Whisper </a> for the same audio 
            file input. Both models use the <a href="https://huggingface.co/distil-whisper/distil-large-v2#long-form-transcription"> chunked long-form transcription algorithm </a> 
            in 🤗 Transformers with Flash Attention support. To ensure fair usage of the Space, we ask that audio 
            file inputs are kept to < 30 mins.
            """
        )
        audio = gr.components.Audio(type="filepath", label="Audio input")
        button = gr.Button("Transcribe")
        with gr.Row():
            distil_runtime = gr.components.Textbox(label="Distil-Whisper Transcription Time (s)")
            runtime = gr.components.Textbox(label="Whisper Transcription Time (s)")
        with gr.Row():
            distil_transcription = gr.components.Textbox(label="Distil-Whisper Transcription", show_copy_button=True)
            transcription = gr.components.Textbox(label="Whisper Transcription", show_copy_button=True)
        button.click(
            fn=transcribe,
            inputs=audio,
            outputs=[distil_transcription, distil_runtime, transcription, runtime],
        )
    demo.queue(max_size=10).launch()