Making automatic speech recognition work on large files with Wav2Vec2 in 🤗 Transformers
Hugging Face Blog
Hugging Face implemented a technique using Wav2Vec2's Connectionist Temporal Classification architecture to enable automatic speech recognition on arbitrarily long audio files by processing overlapping chunks and discarding predictions from the chunk edges. The most popular Wav2Vec2 checkpoint on Hugging Face Hub receives over 250,000 monthly downloads, and users can now process hour-long files by adding a `chunk_length_s` parameter to the pipeline. This approach also enables live speech recognition where transcriptions appear as users speak, without requiring users to wait for the full audio to be processed.