OpenAI Released Two Audio Transcription Models
OpenAI introduced two speech recognition models in the API. GPT-Live-Transcribe is designed for low-latency real-time transcription, while GPT-Transcribe is intended for processing pre-recorded audio files and batch jobs.
The models are better at recognizing different languages and accents, short phrases, numbers, professional terminology, and speech amid heavy background noise. Developers can provide the system with a description of the recording, keywords, expected languages, and previous utterances.
Additional context improved GPT-Transcribe’s semantic accuracy in the Context Aware ASR test from 41.6% to 45.2%. On the multilingual Common Voice dataset, the model achieved a transcription error rate of 19.27%.
Why it matters
- —Developers now have dedicated models for live speech recognition and processing pre-recorded audio.
- —Providing context helps the models recognize names, numbers, and specialized terms more accurately.
- —Support for multiple languages, accents, and noisy recordings broadens the potential applications of voice services.
Key facts
- GPT-Live-Transcribe processes audio in real time and is designed for low latency.
- GPT-Transcribe is optimized for pre-recorded files and batch processing.
- Developers can provide a recording description, keywords, languages, and previous utterances.
- In Context Aware ASR, context improved accuracy from 41.6% to 45.2%.
- On Common Voice, the error rate across 22 languages was 19.27%.
The full text is in the original source. Here we provide a brief summary and key facts.
