from faster_whisper import WhisperModel
import numpy as np
import scipy.signal

model_size = "base.en"

model = WhisperModel(model_size, device="cpu", compute_type="float32")

def process_audio(audio_file):
    sample_rate, audio_data = audio_file

    if audio_data.ndim > 1 and audio_data.shape[1] > 1:
        # Mix stereo channels by averaging them
        audio_data = np.mean(audio_data, axis=1)

    #normalise audio data
    np_audio_float32 = audio_data.astype(np.float32) / 32768.0

    np_audio_16k = scipy.signal.resample(np_audio_float32, int(len(np_audio_float32) * 16000 / sample_rate))
    return np_audio_16k

def transcribe(audio):
    segments, info = model.transcribe(process_audio(audio), beam_size=5, language='en')
    text = "".join([segment.text for segment in segments])
    return text