ovindu-a/sinhala-gemma-asr

Sinhala Gemma ASR

Sinhala speech recognition: a CTC reader trained on top of EmbeddingGemma 2's audio encoder (174 h of Sinhala: SinhaSpeech v6 + OpenSLR-52), decoded with a Sinhala 4-gram KenLM language model. WER 14.7 / CER 3.7 on the SinhaSpeech v6 test set (15,860 utterances).

These are the files the Sinhala Transcribe Mac app downloads on first launch.

• ctc/ the CTC reader (config, unit vocabulary, weights)
• lm/ the 4-gram KenLM (Sinhala Wikipedia, FineWeb-2 and training transcripts; dev/test sentences removed)
• encoder/ EmbeddingGemma 2's audio tower only (bf16, as loaded by transformers) with its config and tokenizer, from google/embeddinggemma-2@914f7f89142e33e77833254d9c9b90c3cef7303b

Licence

The encoder weights and the CTC reader trained on top of them are Gemma model derivatives, provided under and subject to the Gemma Terms of Use (https://ai.google.dev/gemma/terms) and the Gemma Prohibited Use Policy (https://ai.google.dev/gemma/prohibited_use_policy). Not affiliated with or endorsed by Google.