Anyway to use hotwords for parakeet v2?

[Michael-1001]

Sorry will try to make the inquiry quick because perhaps it's an obvious no (or yes?)

But I know NVIDIA's framework allows for phrase boosting (an example of some discussion here: https://github.com/NVIDIA-NeMo/NeMo/issues/14500) but beam search doesn't work on sherpa-onnx-nemo-parakeet-tdt-0.6b-v2-fp16 and thus cannot use hotwords (also apparently theirs you can use greedy with keyphrases too, and apparently has optional bpe which I think is what hotwords in sherpa does).

Is there perhaps some alternative way that I may not know about? For my purpose even re-weighting tokens in decoding is helpful. My apologies if I've made a misunderstanding as I'm somewhat still new to all of this and may not understand the correct.

[dosubot[bot]]

Hi @Michael-1001! I'm Dosu and I’m helping the sherpa-onnx team.

Hotwords (phrase boosting) in sherpa-onnx are only supported for transducer models when using the modified_beam_search decoding method—other models and decoding methods, including greedy_search, do not support hotwords. There’s no alternative like token re-weighting or phrase boosting for Parakeet v2 models in sherpa-onnx if you can’t use modified_beam_search with them. Open vocabulary keyword spotting is a separate feature and doesn’t provide general phrase boosting for ASR tasks either. If Parakeet v2 isn’t recognized as a transducer model by sherpa-onnx or doesn’t support modified_beam_search, hotword boosting isn’t available at this time (source).

To reply, just mention @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}   

[Michael-1001]

Damn for a bot that's a decent response.

I guess I'm just leaving it open anyhow in case there are any other ideas someone may have. I could just use their library but I'm guessing the performance would be a bit questionable on a CPU...