Add Welsh Parakeet TDT-CTC 110M model (stt_cy_parakeet_tdt_ctc_110m)
Adds models/stt_cy_parakeet_tdt_ctc_110m.tar.gz, a Welsh (cy) acoustic model for
Speech-to-Phrase, from Uned Technolegau Iaith (techiaith), Bangor University.
- Same architecture, ONNX export script (k2-fsa/sherpa-onnx) and package layout as
stt_en_parakeet_tdt_ctc_110m:stt_cy_parakeet_tdt_ctc_110m/holdingmodel.onnx,tokens.txtandMODEL_CARD.md, plustokens.json,LICENSEandNOTICE. - Source: https://huggingface.co/techiaith/stt_cy_parakeet_tdt_ctc_110m (commit
0f7d1f331f). sha2568215c7d3d2176e430a1e53d743adf3e0da896eaaadd4d398cf0510e3e57bc2f8. - Licence Apache 2.0. It is a fine-tune of
nvidia/parakeet-tdt_ctc-110m(© NVIDIA,
CC BY 4.0), whose terms still apply to that part. The package'sNOTICEcarries NVIDIA's
attribution and those of the training data (FLEURS and four room impulse response corpora,
all CC BY 4.0); Apache section 4 asks that it stays with every copy. - Needs
MODEL_MAX_SCORE["stt_cy_parakeet_tdt_ctc_110m"] = 4.0; the NeMo default of
5.0 would accept everyday speech.
On held-out recordings of real Welsh speakers (four speakers, close and arm's-length
microphones, 328 takes), Speech-to-Phrase got 197 of 200 on-device commands right and
rejected all 60 everyday sentences. On a Home Assistant Green it transcribes a command
in about 2 s.
Related: the Welsh speech_to_phrase sentences are in OHF-Voice/intents#4286. The
OHF-Voice/apps pull request that registers the model and moves HF_REVISION to the
commit that adds this file will follow once this is merged.
Updated: the model is now released under Apache 2.0, following Bangor University's model licensing policy. The weights are unchanged; the package adds NOTICE, which carries NVIDIA's CC BY 4.0 attribution for the base model and the training data's attributions. New sha256: 8215c7d3d2176e430a1e53d743adf3e0da896eaaadd4d398cf0510e3e57bc2f8.