In September 2022, OpenAI open-sourced Whisper, an encoder-decoder Transformer speech model trained on 680,000 hours of multilingual and multitask supervised audio data, achieving near-human robustness across accents and background noise.

Weakly Supervised Scaling Across 96 Languages

On September 21, 2022, OpenAI released the paper and open-source weights for **Whisper** ('Robust Speech Recognition via Large-Scale Weak Supervision'). Unlike previous automatic speech recognition (ASR) models trained on small, pristine studio audio datasets, Whisper was trained on 680,000 hours of diverse web audio—of which 117,000 hours cover 96 non-English languages and 125,000 hours consist of `X -> en` speech-to-text translation data. Using an encoder-decoder Transformer architecture fed with 80-channel log-Mel spectrograms, Whisper handles transcription, voice activity detection, language identification, and speech translation in a single model.

Transforming Subtitling Time-Coding and Transcription Workflows

For the audiovisual localization and transcription sector, Whisper's robustness against background noise, technical jargon, and regional accents marks a major leap forward in generating initial time-coded transcripts (SRT/VTT). However, OpenAI's technical report also documents that sequence-to-sequence ASR models can occasionally exhibit 'hallucinations' (repeating phrases during silent audio segments or paraphrasing fast speech)—underscoring why professional human transcriptionists and subtitlers must review and verify every time-coded segment for legal, medical, and broadcast deliverables.

Avete bisogno di traduzioni multilingue certificate, asseverazioni legali o localizzazione conforme agli standard ISO per i vostri progetti internazionali? Richiedete un preventivo gratuito personalizzato ad Alphatrad e riceverete un piano operativo dettagliato entro 24 ore.