OpenAI Launches Whisper API to Bring Advanced Speech-to-Text to Developers
OpenAI releases a hosted Whisper API alongside its ChatGPT API, offering fast and affordable transcription and translation. The system excels at handling accents and background noise but faces challenges with word prediction and language biases.
OpenAI launches the Whisper API, a hosted version of its open source speech-to-text model, to accompany the recent rollout of the ChatGPT API. Priced at $0.006 per minute, this automatic speech recognition system provides robust transcription and translation into English across multiple languages. It supports a wide variety of audio formats, including M4A, MP3, MP4, WAV, and WEBM, making it highly accessible for enterprise integration.
The Whisper API stands out from competing tech giant systems because its underlying model trains on 680,000 hours of diverse, multitask web data. OpenAI president Greg Brockman explains that this extensive training improves recognition of unique accents, background noise, and technical jargon. Furthermore, OpenAI optimizes the hosted API to be significantly faster and more convenient for developers than running the open source model independently.
Despite its strengths, the system exhibits notable limitations that developers must consider. OpenAI warns that Whisper occasionally adds words to transcriptions that no one actually speaks due to its next-word prediction behavior. Additionally, the system struggles with a higher error rate for languages underrepresented in its training data, highlighting a persistent bias issue that continues to plague the broader speech recognition industry.