using the large model within python

[bradleyb127]

Installed Whisper and everything works from the command line and within a python script.

However, when using the following command line command, I get much better results (as expected):
whisper --model large ".\20230428.mp3"

There are words in the audio that are transcribed correctly this way.

However, I'm not having any luck using the large model from within a python script.

I have tried the following:

import os
import openai
openai.api_key = "key"
audio_file = open("20230428.mp3", "rb")
transcript = openai.Audio.transcribe("large", audio_file)
print(transcript.text)

and:

import whisper

model = whisper.load_model("large-v2")
result = model.transcribe("20230428.mp3")
print (result["text"])

and:

import whisper
load the model
model = whisper.load_model("large-v2")
get transcription
result = model.transcribe("20230428.mp3", 
language="english")
result contains 3 output,
result['text'] --> complete transcription that with punctuations
result['segments'] --> segment wise transcription with timestamps and other details 
result['langauge'] --> detected language of the audio
print (result["text"])

I'm able to get the file transcribed, but not with the accuracy as the cli results and it runs much faster so for those two reasons, I don't believe it's actually using the large (or large-v2) model.

Any suggestions would be appreciated as this has been a wall I can't figure out for about a week now.

And, thank you.

Brad

[glangford]

This is likely because the command line invocation is using different default settings for the transcription, as described here:

• Much slower via command line tool than in Python? #177 (comment)

You can just load the "large" model, it is the same as "large-v2". Try changing the transcription line in your 3rd code block to:

model.transcribe("20230428.mp3", beam_size=5, best_of=5)

[bradleyb127]

That worked - thank you! I was able to get the same results within a python script as I was from the command line by changing the transcription line as you suggested.