Skip to main content

Hello,

I am using the whisper-large-v3 model in Playground and getting great results on one of my audio files. However, when I use the same model through the API, the results are significantly worse. The Turbo model works well, providing the same high-quality results as in Playground.

My main question is why the results in Playground and the API don't match. Here is my code...

import os

 

from groq import Groq

 

client = Groq()

 

file_path = "/content/denoised (29).m4a"

with open(file_path, "rb") as file:

transcription = client.audio.transcriptions.create(

file=file,

model="whisper-large-v3",

# prompt="Specify context or spelling",

response_format="verbose_json",

timestamp_granularities=a"word", "segment"],

language="uk",

temperature=0.0

)

transcription.to_json()

Could you help me understand the reasons for this discrepancy in quality and suggest ways to improve the results when using the API?

We're here to help
Reply directly to this email or through our Messenger
 
e68589054fce96b5e3ad54aabd6c0d23.png Groq
 
 
 
Powered by Intercom
intercom
 

Hi Petr,

Under the hood, the Playground uses the same API calls as the console docs.

Your code looks correct though, so I’ll reach out to you through email so I can replicate your issue with the same audio file. Is the audio file in Ukrainian?

Best,
Jan

Reply


Terms & ConditionsCookie settings