How dataoceanai/dolphin works
Project type and purpose
• Dolphin is a Python-based command-line tool and library providing multilingual, multitask automatic speech recognition (ASR).
• From the README: it performs speech recognition, voice activity detection (VAD), segmentation, and language identification (LID) over 40 Eastern languages and 22 Chinese dialects.
File structure
• Top-level files:
• setup.py, requirements.txt – packaging and dependencies
• README.md, LICENSE – documentation
• dolphin/ – the library source
• figures/, languages.md – extra docs/assets
• Inside dolphin/:
• __main__.py – CLI entrypoint
• audio.py – audio loading & waveform conversion
• transcribe.py – the high-level pipeline orchestration
• model.py – model loading & forward pass (encoder-decoder + joint CTC-Attention)
• beam_search.py – decoding logic
• scorefilter.py – post-processing, filtering transcriptions
• constants.py, languages.py, version.py – config, supported languages, metadata
• assets/ – pre-trained BPE model, feature stats, YAML config
README for architectural hints
• Dolphin largely follows the Whisper/OWSM design:
• Encoder: E-Branchformer
• Decoder: standard Transformer
• Joint CTC-Attention architecture
• Two-level language tokens (language + region)
• Dependencies: FFmpeg (external binary), PyTorch (implied), standard Python libs