Configuration Parsing Warning:In adapter_config.json: "peft.task_type" must be a string
Konkani Whisper STT
Latest rank-64 adapter trained for one epoch from the base checkpoint on corrected Romi Konkani synthetic speech. The previous, lower-quality adapter has been removed from Hugging Face; this repository now contains the latest trained model.
Training clips: 18,825. Completed optimizer steps: 4,707. Validation and test examples were not included in training. Single synthetic voice; machine-corrected transcripts. Load alongside the base model; base weights are not duplicated here.
Final validation: WER 40.95% / CER 9.53% on 128 held-out synthetic single-voice clips. These are not natural-speaker validation metrics. Transcripts use plain Romi Konkani without vocal tags. Whisper uses its Marathi language token as a proxy.
Dataset: Konkani synthetic speech,
pinned to 72e1c29c1f47040d5b192b5c97c75f646aa45f80. Source/provider terms and synthetic-data
provenance remain documented in the dataset card. Training metrics and
generation/transcription examples are included.
- Downloads last month
- 42