OWSM CTC v4 1B

OWSM (Open Whisper-style Speech Model) is a community effort led by CMU's WAVLab and the ESPnet team to build fully reproducible, openly trained alternatives to OpenAI Whisper. The v4 CTC variant is an encoder-only model using hierarchical multi-task self-conditioned CTC with an E-Branchformer encoder.

• ~1B parameter E-Branchformer encoder with CTC heads (no autoregressive decoder)
• 128 Mel filterbanks, 8× subsampling → 80 ms frame resolution
• Trained on 16 kHz audio in fixed 30 s chunks

• Fully open: weights, training data (curated YODAS), code, and logs are all publicly released
• Faster than AED models: several× faster inference than Whisper/Canary thanks to CTC greedy decoding
• Supports ASR, any-to-any speech translation, language identification, and utterance-level timestamps
• Won Best Student Paper at INTERSPEECH 2025 (OWSM v4)

Reproducible research, multilingual transcription/translation across 75 languages, forced alignment (CTC segmentation), and as a base for further fine-tuning where training transparency is required.