app.py

import spaces
import torch

import gradio as gr
import yt_dlp as youtube_dl
from transformers import pipeline
from transformers.pipelines.audio_utils import ffmpeg_read

import tempfile
import os
import time

# Available models to choose from
MODEL_OPTIONS = ["BUT-FIT/DeCRED-base", "BUT-FIT/DeCRED-small", "BUT-FIT/ED-base", "BUT-FIT/ED-small"]
DEFAULT_MODEL = MODEL_OPTIONS[1]

BATCH_SIZE = 8
FILE_LIMIT_MB = 1000
YT_LENGTH_LIMIT_S = 3600  # limit to 1 hour YouTube files

device = 0 if torch.cuda.is_available() else "cpu"


# Function to initialize pipeline based on model selection
def initialize_pipeline(model_name):
    pipe = pipeline(
        task="automatic-speech-recognition",
        model=model_name,
        feature_extractor=model_name,
        chunk_length_s=30,
        device=device,
        trust_remote_code=True
    )
    pipe.type = "seq2seq"
    return pipe


# Initialize the pipeline with a default model (it will be updated after user selects one)
pipe = initialize_pipeline(DEFAULT_MODEL)
pipe.type = "seq2seq"


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

    # Update the pipeline with the selected model
    pipe = initialize_pipeline(selected_model)

    text = pipe(inputs, batch_size=BATCH_SIZE)["text"]
    return text


def _return_yt_html_embed(yt_url):
    video_id = yt_url.split("?v=")[-1]
    HTML_str = (
        f'<center> <iframe width="500" height="320" src="https://www.youtube.com/embed/{video_id}"> </iframe>'
        " </center>"
    )
    return HTML_str


def download_yt_audio(yt_url, filename):
    info_loader = youtube_dl.YoutubeDL()

    try:
        info = info_loader.extract_info(yt_url, download=False)
    except youtube_dl.utils.DownloadError as err:
        raise gr.Error(str(err))

    file_length = info["duration_string"]
    file_h_m_s = file_length.split(":")
    file_h_m_s = [int(sub_length) for sub_length in file_h_m_s]

    if len(file_h_m_s) == 1:
        file_h_m_s.insert(0, 0)
    if len(file_h_m_s) == 2:
        file_h_m_s.insert(0, 0)
    file_length_s = file_h_m_s[0] * 3600 + file_h_m_s[1] * 60 + file_h_m_s[2]

    if file_length_s > YT_LENGTH_LIMIT_S:
        yt_length_limit_hms = time.strftime("%HH:%MM:%SS", time.gmtime(YT_LENGTH_LIMIT_S))
        file_length_hms = time.strftime("%HH:%MM:%SS", time.gmtime(file_length_s))
        raise gr.Error(f"Maximum YouTube length is {yt_length_limit_hms}, got {file_length_hms} YouTube video.")

    ydl_opts = {"outtmpl": filename, "format": "worstvideo[ext=mp4]+bestaudio[ext=m4a]/best[ext=mp4]/best"}

    with youtube_dl.YoutubeDL(ydl_opts) as ydl:
        try:
            ydl.download([yt_url])
        except youtube_dl.utils.ExtractorError as err:
            raise gr.Error(str(err))


@spaces.GPU
def yt_transcribe(yt_url, selected_model, max_filesize=75.0):
    html_embed_str = _return_yt_html_embed(yt_url)

    # Update the pipeline with the selected model
    pipe = initialize_pipeline(selected_model)

    with tempfile.TemporaryDirectory() as tmpdirname:
        filepath = os.path.join(tmpdirname, "video.mp4")
        download_yt_audio(yt_url, filepath)
        with open(filepath, "rb") as f:
            inputs = f.read()

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

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

    return html_embed_str, text


demo = gr.Blocks(theme=gr.themes.Ocean())

mf_transcribe = gr.Interface(
    fn=transcribe,
    inputs=[
        gr.Audio(sources="microphone", type="filepath"),
        gr.Dropdown(choices=MODEL_OPTIONS, label="Model", value=DEFAULT_MODEL)
    ],
    outputs="text",
    title="Transcribe Audio",
    description=(
        "Transcribe long-form microphone or audio inputs with the click of a button! Select a model from the dropdown."
    ),
    allow_flagging="never",
)

file_transcribe = gr.Interface(
    fn=transcribe,
    inputs=[
        gr.Audio(sources="upload", type="filepath", label="Audio file"),
        gr.Dropdown(choices=MODEL_OPTIONS, label="Model", value=DEFAULT_MODEL)
    ],
    outputs="text",
    title="Transcribe Audio",
    description=(
        "Transcribe audio files with the click of a button! Select a model from the dropdown."
    ),
    allow_flagging="never",
)

yt_transcribe = gr.Interface(
    fn=yt_transcribe,
    inputs=[
        gr.Textbox(lines=1, placeholder="Paste the URL to a YouTube video here", label="YouTube URL"),
        gr.Dropdown(choices=MODEL_OPTIONS, label="Model", value=DEFAULT_MODEL)
    ],
    outputs=["html", "text"],
    title="Transcribe YouTube",
    description=(
        """
        ### *Currently only works on local instances of this space, as youtube-dl does not function from Hugging Face servers.*
        Transcribe long-form YouTube videos with the click of a button! Select a model from the dropdown."""
    ),
    allow_flagging="never",
)

with demo:
    gr.TabbedInterface([mf_transcribe, file_transcribe, yt_transcribe], ["Microphone", "Audio file", "YouTube"])

    gr.Markdown(
        """
        ## Overview
        This space demonstrates the performance of **DeCRED** (**De**coder-**C**entric **R**egularization in **E**ncoder-**D**ecoder) for automatic speech recognition (ASR). 
        DeCRED enhances model robustness and generalization, particularly in out-of-domain scenarios, by introducing auxiliary classifiers in the decoder layers of encoder-decoder ASR architectures.
        
        ## Key Features
        - **Auxiliary Classifiers**: DeCRED integrates auxiliary classifiers in the decoder module to regularize training, improving the model’s ability to generalize across domains.
        - **Enhanced Decoding**: It proposes two new decoding strategies that leverage auxiliary classifiers to re-estimate token probabilities, resulting in more accurate ASR predictions.
        - **Strong Baseline**: Built on the **E-branchformer** architecture, DeCRED achieves competitive word error rates (WER) compared to Whisper-medium and OWSM v3 while requiring significantly less training data and a smaller model size.
        - **Out-of-Domain Performance**: DeCRED demonstrates strong generalization, reducing WERs by 2.7 and 2.9 points on the AMI and Gigaspeech datasets, respectively.
        
        ## Disclaimer
        This space currently runs on basic CPU hardware, so generation might take a bit longer (approximately four times the length of the audio). 
        You can clone the repository and run it locally for better performance. 
        Please refer to the [Hugging Face documentation](https://huggingface.co/docs/hub/spaces-overview#clone-the-repository) 
        for instructions on how to clone the repository and run it locally. 
        The model is not perfect and may make errors, so please use it responsibly.
        
        ## Explore the Models
        - [DeCRED Base](https://huggingface.co/BUT-FIT/DeCRED-base)
        - [DeCRED Small](https://huggingface.co/BUT-FIT/DeCRED-small)
        - [ED Base](https://huggingface.co/BUT-FIT/ED-base)
        - [ED Small](https://huggingface.co/BUT-FIT/ED-small)
        
        ## Citation
        If you use DeCRED in your research, please cite the following paper:
        
        ```bibtex
        @misc{polok2024improvingautomaticspeechrecognition,
              title={Improving Automatic Speech Recognition with Decoder-Centric Regularisation in Encoder-Decoder Models}, 
              author={Alexander Polok and Santosh Kesiraju and Karel Beneš and Lukáš Burget and Jan Černocký},
              year={2024},
              eprint={2410.17437},
              archivePrefix={arXiv},
              primaryClass={eess.AS},
              url={https://arxiv.org/abs/2410.17437}, 
        }
        ```
        """
    )

demo.queue().launch(ssr_mode=False)