Transcribe and return raw text

faster-whisper returns transcription as an iterator of segments, not a single string. Each segment has a start time, end time, and text. The simplest output is the concatenated text of every segment, but you can also keep timestamps for captioning or search.

backend/transcription.py

python

def transcribe(self, audio_file):
    print('Transcribing...')

    segments, info = self.whisper.transcribe(
        audio_file,
        beam_size=5,
        language='en',
        condition_on_previous_text=False,
    )

    text = ' '.join([segment.text for segment in segments]).strip()
    print(f'Raw: {text}')
    return text

beam_size=5 trades a little speed for better accuracy. condition_on_previous_text=False stops the model from reusing earlier context, which prevents hallucination loops on silent audio.

Beam search keeps the top N candidate transcriptions at each step and picks the best sequence overall. beam_size=1 is greedy decoding, fast but lower quality. beam_size=5 is a common sweet spot, costing a bit more compute for a noticeable accuracy bump.

Quiz: Quiz

Loading practice…

Checkpoint: Local Whisper checkpoint

Loading practice…