app.py

#%%
from huggingface_hub import login
from transformers import pipeline
from datasets import load_dataset
import gradio as gr
import os

login(token=os.environ['hf_token'])

atco2 = load_dataset('jlvdoorn/atco2-asr', split='validation')
atcosim = load_dataset('jlvdoorn/atcosim', split='validation')

examples = [atco2[0]['audio'], atcosim[0]['audio'], atco2[1]['audio'], atcosim[1]['audio'], atco2[2]['audio'], atcosim[2]['audio']]
examples_labels = ['Example ' + str(i) for i in len(examples)]

## Try to load a local model if available
# try:
#     whisper = pipeline(model='/mnt/projects/whisper/WhisperANSP/Models/whisper-large-v2-atco2-asr-atcosim-ANSP-3h1m', task='automatic-speech-recognition')
#     ttl = 'Whisper Large v2 - ATCO2-ATCOSIM-ANSP'
#     dis = 'This demo will transcribe ATC audio files by using the Whisper Large v2 model fine-tuned on the ATCO2, ATCOSIM and ANSP datasets. \n \n Further it uses a Named Entity Recognition model to extract callsigns, commands and values from the transcription. \n This model is based on Google\'s BERT model and fine-tuned on the ATCO2 dataset.'

# except:
# whisper = pipeline(model='jlvdoorn/whisper-large-v2-atco2-asr-atcosim')
# ttl = 'Whisper Large v2 - ATCO2-ATCOSIM'
# dis = 'This demo will transcribe ATC audio files by using the Whisper Large v2 model fine-tuned on the ATCO2 and ATCOSIM datasets. \n \n Further it uses a Named Entity Recognition model to extract callsigns, commands and values from the transcription. \n This model is based on Google\'s BERT model and fine-tuned on the ATCO2 dataset.'

bert_atco_ner = pipeline(model='Jzuluaga/bert-base-ner-atc-en-atco2-1h')

whisper_v2 = pipeline(model='jlvdoorn/whisper-large-v2-atco2-asr-atcosim')
whisper_v3 = pipeline(model='jlvdoorn/whisper-large-v3-atco2-asr-atcosim')
#%%
def transcribe(audio_file, audio_mic, model_version):
    if model_version == 'large-v2':
        whisper = whisper_v2
        ttl = 'Whisper Large v2 - ATCO2-ATCOSIM'
        dis = 'This demo will transcribe ATC audio files by using the Whisper Large v2 model fine-tuned on the ATCO2 and ATCOSIM datasets. \n \n Further it uses a Named Entity Recognition model to extract callsigns, commands and values from the transcription. \n This model is based on Google\'s BERT model and fine-tuned on the ATCO2 dataset.'
    elif model_version == 'large-v3':
        whisper = whisper_v3
        ttl = 'Whisper Large v3 - ATCO2-ATCOSIM'
        dis = 'This demo will transcribe ATC audio files by using the Whisper Large v3 model fine-tuned on the ATCO2 and ATCOSIM datasets. \n \n Further it uses a Named Entity Recognition model to extract callsigns, commands and values from the transcription. \n This model is based on Google\'s BERT model and fine-tuned on the ATCO2 dataset.'
    if audio_mic is not None:
        return whisper(audio_mic)['text']
    elif audio_file is not None:
        return whisper(audio_file)['text']
    else:
        return 'There was no audio to transcribe...'

#%%
def extractCallSignCommand(transcription):
    if type(transcription) is str:
        result = bert_atco_ner(transcription)
        callsigns = []
        commands = []
        values = []
        for item in result:
            if 'callsign' in item['entity']:
                callsigns.append(item['word'])
            if 'command' in item['entity']:
                commands.append(item['word'])
            if 'value' in item['entity']:
                values.append(item['word'])
                
        return 'Callsigns: ' + ', '.join(callsigns) + '\nCommands: ' + ', '.join(commands) + '\nValues: ' + ', '.join(values)
    else:
        return 'There was no transcription to extract a callsign or command from...'

#%%
def transcribeAndExtract(audio_file, audio_mic, transcribe_only, model_version):
    transcription = transcribe(audio_file, audio_mic, model_version)
    if not transcribe_only:
        callSignCommandValues = extractCallSignCommand(transcription)
    else:
        callSignCommandValues = ''
    return transcription, callSignCommandValues

#%%
iface = gr.Interface(
        fn=transcribeAndExtract,
        inputs=[
            gr.Audio(source='upload', type='filepath', interactive=True), 
            gr.Audio(source='microphone', type='filepath'), 
            
            gr.Checkbox(label='Transcribe only', default=False),
            gr.Dropdown(choices=['large-v2', 'large-v3'], value='large-v3', label='Whisper model version'),
            ],
        outputs=[gr.Text(label='Transcription'), gr.Text(label='Callsigns, commands and values')],
        title='Whisper ATC - Large v3',
        description='Transcribe and extract',
        examples = examples,
)

#%%
#iface.launch(server_name='0.0.0.0', server_port=9000)
iface.launch()