Groq Whisper API: An Innovative Communication Interface

In the midst of our exploration of Claude 3.5 Sonnet, GroqInc unveiled an API for OpenAI’s latest Whisper-large-v3 model. This solution converts speech to text at an incredibly rapid pace. Moreover, it’s offered at the affordable rate of $0.03 per transcription hour! For those interested in leveraging Groq’s Whisper API alongside the OpenAI client, here’s how you can do it:

from openai import OpenAI
groq = OpenAI(api_key="YOUR_GROQ_API_KEY",
base_url="
audio_file = open("/path/to/your_audio.mp3", "rb")
transcript = groq.audio.transcriptions.create(
model="whisper-large-v3", file=audio_file, response_format="text"
print(transcript)

Understanding Whisper

Whisper is an advanced model specifically developed for automatic speech recognition (ASR) and speech-to-text translation. Its training was extensive, involving 680,000 hours of labeled data, which has equipped it with the capability to adapt to a variety of datasets and applications without bespoke adjustments.

This cutting-edge model was first introduced through the study “Robust Speech Recognition via Large-Scale Weak Supervision” by Alec Radford and others from OpenAI. The model’s source code is openly accessible for reference.

Key features of the Whisper large-v3 iteration include:

• An upgrade to 128 Mel frequency bins for its input, up from the previous 80.
• The addition of a language token specifically for Cantonese.

The development of Whisper large-v3 involved training on 1 million hours of weakly labeled audio alongside an additional 4 million hours of pseudolabeled audio content, originally recognized by the Whisper large-v2 model. This comprehensive training covered 2.0 epochs across the combined datasets.

Among its advancements, the large-v3 model boasts a notable reduction in recognition errors, ranging from 10% to 20%, compared to its predecessor, large-v2.

Learn more: Comparing Whisper large-v3 and large-v2 models

Related readings: