DEV Community

El Bruno for Microsoft Azure

Posted on • Originally published at on

#AI – Using OpenAI Whisper to convert audio to text, in Python, from my podcast episode (in Spanish!)


Today’s post is how to use Whisper to get the text transcription from my podcast episode audio. Based on the Podcast Copilot sample, I decided to use Whisper to do this. Let’s start!

OpenAI – Whisper

Whisper is a general-purpose speech recognition model. It is trained on a large dataset of diverse audio and is also a multitasking model that can perform multilingual speech recognition, speech translation, and language identification.

Whisper GitHub

You can get more information about Whisper here

And, once installed (simple python package), it’s super easy to use:

import whisper

model = whisper.load_model("base")
result = model. Transcribe("audio.mp3")

Enter fullscreen mode Exit fullscreen mode

Super easy! However, this approach may trigger some memory problems with large audio files.

The approach in the original Podcast Copilot is to split the podcast audio into small chunks and get the transcript for each one of the small pieces. In example:

# Chunk up the audio file 
sound_file = AudioSegment.from_mp3(podcast_audio_file)
audio_chunks = split_on_silence(sound_file, min_silence_len=1000, silence_thresh=-40 )
count = len(audio_chunks)
print("Audio split into " + str(count) + " audio chunks \n")

# Call Whisper to transcribe audio
model = whisper.load_model("base")
transcript = ""
for i, chunk in enumerate(audio_chunks):    
    if i < 10 or i > count - 10:
        out_file = "chunk{0}.wav".format(i)
        print("\r\nExporting >>", out_file, " - ", i, "/", count)
        chunk.export(out_file, format="wav")
        result = model.transcribe(out_file)
        transcriptChunk = result["text"]

        transcript += " " + transcriptChunk

# Print transcript
print("Transcript: \n")

Enter fullscreen mode Exit fullscreen mode

And that’s it! Depending on your machine, this may take some time. I added a start/end check, so you can get an idea of how much time you need to spend processing an episode.

In example, for a 10 min audio, transcript done with an estimated time of 06:05.

transcript done for the last chuck with an estimated time of 06:05.

The full source code is here:

Happy coding!


El Bruno

More posts in my blog

More info in

Top comments (0)