app.py

import os
# os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "expandable_segments:True"

import gc
from functools import partial
import gradio as gr
import torch
from speechbrain.inference.interfaces import Pretrained, foreign_class
from transformers import T5Tokenizer, T5ForConditionalGeneration
import librosa
import whisper_timestamped as whisper
from transformers import AutoModelForSpeechSeq2Seq, AutoProcessor, pipeline, Wav2Vec2ForCTC, AutoProcessor
    

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
torch.backends.cuda.matmul.allow_tf32 = True


def clean_up_memory():
    gc.collect()
    torch.cuda.empty_cache()


def recap_sentence(string):
    # Restore capitalization and punctuation using the model
    inputs = recap_tokenizer(["restore capitalization and punctuation: " + string], return_tensors="pt", padding=True).to(device)
    outputs = recap_model.generate(**inputs, max_length=768, num_beams=5, early_stopping=True).squeeze(0)
    recap_result = recap_tokenizer.decode(outputs, skip_special_tokens=True)
    return recap_result


def return_prediction_w2v2(mic=None, file=None, device=device):
    if mic is not None:
        waveform, sr = librosa.load(mic, sr=16000)
        waveform = waveform[:30*sr]
        w2v2_result = w2v2_classifier.classify_file_w2v2(waveform, device)
    elif file is not None:
        waveform, sr = librosa.load(file, sr=16000)
        waveform = waveform[:30*sr]
        w2v2_result = w2v2_classifier.classify_file_w2v2(waveform, device)
    else:
        return "You must either provide a mic recording or a file"

    recap_result = recap_sentence(w2v2_result[0])

    # If the letter after punct is small, recap it
    for i, letter in enumerate(recap_result):
        if i > 1 and recap_result[i-2] in [".", "!", "?"] and letter.islower():
            recap_result = recap_result[:i] + letter.upper() + recap_result[i+1:]

    clean_up_memory()
    return recap_result



def return_prediction_whisper(mic=None, file=None, device=device):
    if mic is not None:
        waveform, sr = librosa.load(mic, sr=16000)
        waveform = waveform[:30*sr]
        whisper_result = whisper_classifier.classify_file_whisper_mkd(waveform, device)
    elif file is not None:
        waveform, sr = librosa.load(file, sr=16000)
        waveform = waveform[:30*sr]
        whisper_result = whisper_classifier.classify_file_whisper_mkd(waveform, device)
    else:
        return "You must either provide a mic recording or a file"

    recap_result = recap_sentence(whisper_result[0])

    # If the letter after punct is small, recap it
    for i, letter in enumerate(recap_result):
        if i > 1 and recap_result[i-2] in [".", "!", "?"] and letter.islower():
            recap_result = recap_result[:i] + letter.upper() + recap_result[i+1:]

    clean_up_memory()
    return recap_result


def return_prediction_compare(mic=None, file=None, device=device):
    # pipe_whisper.model.to(device)
    # mms_model.to(device)
    if mic is not None:
        waveform, sr = librosa.load(mic, sr=16000)
        waveform = waveform[:30*sr]
        whisper_mkd_result = whisper_classifier.classify_file_whisper_mkd(waveform, device)
        # result_generator_w2v2 = w2v2_classifier.classify_file_w2v2(mic, device)
        whisper_result = whisper_classifier.classify_file_whisper(waveform, pipe_whisper, device)
        mms_result_generator = whisper_classifier.classify_file_mms(waveform, processor_mms, mms_model, device)
        
    elif file is not None:
        waveform, sr = librosa.load(file, sr=16000)
        waveform = waveform[:30*sr]
        whisper_mkd_result = whisper_classifier.classify_file_whisper_mkd(waveform, device)
        # result_generator_w2v2 = w2v2_classifier.classify_file_w2v2(file, device)
        whisper_result = whisper_classifier.classify_file_whisper(waveform, pipe_whisper, device)
        mms_result_generator = whisper_classifier.classify_file_mms(waveform, processor_mms, mms_model, device)
    else:
        return "You must either provide a mic recording or a file"
    # pipe_whisper.model.to("cpu")
    # mms_model.to("cpu")

    segment_results_whisper = ""
    prev_segment_whisper = ""
    # segment_results_w2v2 = ""
    # prev_segment_w2v2 = ""
    segment_results_mms = ""
    prev_segment_mms = ""

    recap_result_whisper_mkd = recap_sentence(whisper_mkd_result[0])
    recap_result_whisper = recap_sentence(whisper_result[0])
    recap_result_mms = recap_sentence(mms_result_generator[0])

    # If the letter after punct is small, recap it
    for i, letter in enumerate(recap_result_whisper_mkd):
        if i > 1 and recap_result_whisper_mkd[i-2] in [".", "!", "?"] and letter.islower():
            recap_result_whisper_mkd = recap_result_whisper_mkd[:i] + letter.upper() + recap_result_whisper_mkd[i+1:]

    for i, letter in enumerate(recap_result_whisper):
        if i > 1 and recap_result_whisper[i-2] in [".", "!", "?"] and letter.islower():
            recap_result_whisper = recap_result_whisper[:i] + letter.upper() + recap_result_whisper[i+1:]

    for i, letter in enumerate(recap_result_mms):
        if i > 1 and recap_result_mms[i-2] in [".", "!", "?"] and letter.islower():
            recap_result_mms = recap_result_mms[:i] + letter.upper() + recap_result_mms[i+1:]
    
    clean_up_memory()
    return "Буки-Whisper:\n" + recap_result_whisper_mkd + "\n\n" + "MMS:\n" + recap_result_mms + "\n\n" + "OpenAI Whisper:\n" + recap_result_whisper
        # yield "Our W2v2: \n" + segment_results_w2v2 + "\n\n" + "MMS transcript:\n" + segment_results_mms



# Load Whisper model
model_id = "openai/whisper-large-v3"
whisper_model = AutoModelForSpeechSeq2Seq.from_pretrained(model_id, torch_dtype=torch.float16, low_cpu_mem_usage=True, use_safetensors=True, attn_implementation="sdpa")
whisper_model.to(device)
processor = AutoProcessor.from_pretrained(model_id)
pipe_whisper = pipeline(
    "automatic-speech-recognition",
    model=whisper_model,
    tokenizer=processor.tokenizer,
    feature_extractor=processor.feature_extractor,
    torch_dtype=torch.float16,
    return_timestamps=True,
    device=device,
)


# Load MMS model
model_id = "facebook/mms-1b-all"
processor_mms = AutoProcessor.from_pretrained(model_id)
mms_model = Wav2Vec2ForCTC.from_pretrained(model_id)
mms_model = mms_model.to(device)
mms_model.eval()
processor_mms.tokenizer.set_target_lang("mkd")
mms_model.load_adapter("mkd")



# Create a partial function with the device pre-applied
return_prediction_whisper_with_device = partial(return_prediction_whisper, device=device)
# return_prediction_w2v2_with_device = partial(return_prediction_w2v2, device=device)
return_prediction_with_device_compare = partial(return_prediction_compare, device=device)


# Load the ASR models
whisper_classifier = foreign_class(source="Macedonian-ASR/whisper-large-v3-macedonian-asr", pymodule_file="custom_interface_app.py", classname="ASR")
whisper_classifier = whisper_classifier.to(device)
whisper_classifier.eval()


# Load the T5 tokenizer and model for restoring capitalization
recap_model_name = "Macedonian-ASR/mt5-restore-capitalization-macedonian"
recap_tokenizer = T5Tokenizer.from_pretrained(recap_model_name)
recap_model = T5ForConditionalGeneration.from_pretrained(recap_model_name, torch_dtype=torch.float16)
recap_model.to(device)
recap_model.eval()



mic_transcribe_compare = gr.Interface(
    fn=return_prediction_with_device_compare,
    inputs=gr.Audio(sources="microphone", type="filepath"),
    outputs=gr.Textbox(),
    allow_flagging="never",
    live=False,
)

# file_transcribe_compare = gr.Interface(
#     fn=return_prediction_with_device_compare,
#     inputs=gr.Audio(sources="upload", type="filepath"),
#     outputs=gr.Textbox(),
#     allow_flagging="never",
#     live=False
# )


project_description = '''
## Автори:
1. **Дејан Порјазовски**
2. **Илина Јакимовска**
3. **Ордан Чукалиев**
4. **Никола Стиков**

Оваа колаборација е дел од активностите на **Центарот за напредни интердисциплинарни истражувања ([ЦеНИИс](https://ukim.edu.mk/en/centri/centar-za-napredni-interdisciplinarni-istrazhuvanja-ceniis))** при УКИМ.

## Во тренирањето на овој модел се употребени податоци од:
1. Дигитален архив за етнолошки и антрополошки ресурси ([ДАЕАР](https://iea.pmf.ukim.edu.mk/tabs/view/61f236ed7d95176b747c20566ddbda1a)) при Институтот за етнологија и антропологија, Природно-математички факултет при УКИМ.
2. Аудио верзија на меѓународното списание [„ЕтноАнтропоЗум“](https://etno.pmf.ukim.mk/index.php/eaz/issue/archive) на Институтот за етнологија и антропологија, Природно-математички факултет при УКИМ.
3. Аудио подкастот [„Обични луѓе“](https://obicniluge.mk/episodes/) на Илина Јакимовска
4. Научните видеа од серијалот [„Наука за деца“](http://naukazadeca.mk), фондација [КАНТАРОТ](https://qantarot.substack.com/)
5. Македонска верзија на [Mozilla Common Voice](https://commonvoice.mozilla.org/en/datasets) (верзија 18.0)

## Како да придонесете за подобрување на македонските модели за препознавање на говор?
На  следниот [линк](https://drive.google.com/file/d/18sXbX8OABlMS5uYXfC4y9qlI6YZIJ1fQ/view?usp=sharing) ќе најдете инструкции за тоа како да донирате македонски говор преку платформата Mozilla Common Voice.
'''

# Custom CSS
css = """
.gradio-container {
    background-color: #f0f0f0;  /* Set your desired background color */
}
.custom-markdown p, .custom-markdown li, .custom-markdown h2, .custom-markdown a {
    font-size: 15px !important;
    font-family: Arial, sans-serif !important;
}
.gradio-container {
    background-color: #f3f3f3 !important;
}
"""

transcriber_app = gr.Blocks(css=css, delete_cache=(60, 120))
    
with transcriber_app:
    state = gr.State()
    gr.Markdown(project_description, elem_classes="custom-markdown")

    # gr.TabbedInterface(
    #     [mic_transcribe_whisper, mic_transcribe_compare],
    #     ["Буки-Whisper транскрипција", "Споредба на модели"],
    # )
    # state = gr.State(value=[], delete_callback=lambda v: print("STATE DELETED"))

    gr.TabbedInterface(
        [mic_transcribe_compare],
        ["Споредба на модели"],
    )
    state = gr.State(value=[], delete_callback=lambda v: print("STATE DELETED"))

    transcriber_app.unload(return_prediction_whisper)
    transcriber_app.unload(return_prediction_compare)
    transcriber_app.unload(return_prediction_w2v2)


# transcriber_app.launch(debug=True, share=True, ssl_verify=False)
if __name__ == "__main__":
    transcriber_app.queue()
    transcriber_app.launch(share=True)