app.py

from difflib import Differ

import gradio as gr

import string

import torch
from transformers import (
    AutoModelForSpeechSeq2Seq,
    AutoProcessor,
    pipeline,
)


diction_text = """
How is this leisure to be disposed of? In the public-house? the singing hall? the dancing-saloon?
which hold out seductions somewhat more dangerous, methinks, to honest labour than those presented by a library...
We may well rejoice, then, when we see a room such as this filled with attentive and reflective readers.
"""

diction_script = gr.Textbox(diction_text, interactive=False, show_label=False)

device = "cpu"

torch_dtype = torch.float16 if torch.cuda.is_available() else torch.float32

model_id = "openai/whisper-large-v3"

description = f"""
 <div>
    <p>Welcome to Redmond Barry-oke! </p>
    <p>This app aims to demonstrate the potential of using machine learning to transcribe audio. Users are invited to record themselves reading a brief and abridged excerpt from a speech delivered by Sir Redmond Barry at the opening of The Free Public Library of Ballarat Est in 1869. Once recorded and submitted the app will transcribe and return a "diction" score.</p>
    <p>This app uses {model_id} to perform automated transcription</p>
    <p>A full transcript of Sir Redmond Barry's speech can be read in the <a href="https://latrobejournal.slv.vic.gov.au/latrobejournal/issue/latrobe-26/t1-g-t3.html" target="_blank">La Trobe Journal</a></p>
 </div>
"""


model = AutoModelForSpeechSeq2Seq.from_pretrained(
    model_id, low_cpu_mem_usage=True, use_safetensors=True
)
model.to(device)

processor = AutoProcessor.from_pretrained(model_id)

pipe = pipeline(
    task="automatic-speech-recognition",
    model=model,
    tokenizer=processor.tokenizer,
    feature_extractor=processor.feature_extractor,
    max_new_tokens=128,
    chunk_length_s=30,
    batch_size=8,
    return_timestamps=True,
    torch_dtype=torch_dtype,
    device=device,
)


def prepare_text_for_comparison(text_to_clean: str):

    text_to_clean = text_to_clean.translate(str.maketrans("", "", string.punctuation))

    return text_to_clean.casefold()


def diff_texts(diction_text: str, audio_input: str):
    d = Differ()

    return [
        (token[2:], token[0] if token[0] != "" else None)
        for token in d.compare(diction_text, audio_input)
    ]


def calc_score(diff_texts: list) -> float:

    diff_chars = [char for char in diff_texts if char[1] != " "]
    score = float((len(diff_chars) / len(diff_texts)) * 100)

    score = 100 - score

    return score


def transcribe_audio(diction_text, audio):
    result = pipe(audio)

    cleaned_result = prepare_text_for_comparison(result["text"])
    cleaned_diction_text = prepare_text_for_comparison(diction_text)
    diff_text = diff_texts(cleaned_diction_text, cleaned_result)

    score = calc_score(diff_text)

    formatted_score = f"{str(round(score,3))}%"

    return diction_text, diff_text, formatted_score


transcribed_text = gr.Textbox(label="Transcribed text")


highlighted_results = gr.HighlightedText(
    label="Diff",
    combine_adjacent=True,
    show_legend=True,
    color_map={"+": "green", "-": "red"},
)

score = gr.Textbox("0%", label="Barry-oke score")


input_audio = gr.Audio(
    sources=["microphone"],
    type="filepath",
    waveform_options=gr.WaveformOptions(
        waveform_color="#01C6FF",
        waveform_progress_color="#0066B4",
        skip_length=2,
        show_controls=False,
    ),
)

demo = gr.Interface(
    fn=transcribe_audio,
    inputs=[transcribed_text, diction_script, input_audio],
    outputs=[highlighted_results, score],
    title="Redmond Barry-oke",
    description=description,
)


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