app.py

from transformers import AutoModelForSpeechSeq2Seq, AutoProcessor, pipeline, TextIteratorStreamer
from transformers.utils import is_flash_attn_2_available
from transformers.pipelines.audio_utils import ffmpeg_read
from threading import Thread
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")
streamer = TextIteratorStreamer(processor.tokenizer, skip_special_tokens=True)

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"},
    return_timestamps=True
)
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."
        )

    if audio_length_mins >= 0.5:
        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.copy(), batch_size=BATCH_SIZE)["text"]
        yield distil_text, distil_runtime, 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

    else:
        input_features = processor(inputs, sampling_rate=processor.feature_extractor.sampling_rate, return_tensors="pt").input_features

        # Run the generation in a separate thread, so that we can fetch the generated text in a non-blocking way.
        generation_kwargs = dict(input_features, streamer=streamer, max_new_tokens=128, language="en", task="transcribe")
        thread = Thread(target=distilled_model.generate, kwargs=generation_kwargs)

        thread.start()
        start_time = time.time()
        distil_text = ""
        for generated_text in streamer:
            distil_text += generated_text
            yield distil_text, None, None, None

        distil_runtime = time.time() - start_time
        distil_runtime = round(distil_runtime, 2)
        yield distil_text, distil_runtime, None, None

        thread = Thread(target=model.generate, kwargs=generation_kwargs)

        thread.start()
        start_time = time.time()
        text = ""
        for generated_text in streamer:
            text += generated_text
            yield distil_text, distil_runtime, text, None

        runtime = time.time() - start_time
        runtime = round(runtime, 2)
        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: Speed Comparison
                    </h1>
                  </div>
                </div>
            """
        )
        gr.HTML(
            f"""
            <p><a href="https://huggingface.co/distil-whisper/distil-large-v2"> Distil-Whisper</a> is a distilled variant 
            of the <a href="https://huggingface.co/openai/whisper-large-v2"> Whisper</a> model by OpenAI. Compared to Whisper, 
            Distil-Whisper runs 6x faster with 50% fewer parameters, while performing to within 1% word error rate (WER) on
            out-of-distribution evaluation data.</p>
            
            <p>In this demo, we perform a speed comparison between Whisper and Distil-Whisper in order to test this claim.
            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, as well as Flash Attention. To use Distil-Whisper yourself, check the code examples on the
            <a href="https://github.com/huggingface/distil-whisper#1-usage"> Distil-Whisper repository</a>. To ensure fair 
            usage of the Space, we ask that audio file inputs are kept to < 30 mins.</p>
            """
        )
        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],
        )
        gr.Markdown("## Examples")
        gr.Examples(
            [["./assets/example_1.wav"], ["./assets/example_2.wav"]],
            audio,
            outputs=[distil_transcription, distil_runtime, transcription, runtime],
            fn=transcribe,
            cache_examples=False,
        )
    demo.queue(max_size=10).launch()