Instructions to use nphearum/PsarAI-S2S with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use nphearum/PsarAI-S2S with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("automatic-speech-recognition", model="nphearum/PsarAI-S2S", trust_remote_code=True)# pip install -U transformers accelerate # Load model directly from transformers import AutoModel model = AutoModel.from_pretrained("nphearum/PsarAI-S2S", trust_remote_code=True, device_map="auto") - Notebooks
- Google Colab
- Kaggle
Can this model use a 6-gram KenLM language model?
The default output comes from greedy decoding. For higher accuracy, add a 6-gram KenLM language model with CTC beam search. Could you upload it here?
Exactly, I have upload new lm model by splitting as part from original. For KenLM, a bit improve but not much.
I also apply KenLM with word language model to predict how mask can improve most frequent similar word(misspelling).
Exactly, I have upload new lm model by splitting as part from original. For KenLM, a bit improve but not much.
I also apply KenLM with word language model to predict how mask can improve most frequent similar word(misspelling).
With the optional LM (beam , alpha , beta , tuned on the valid sets):What parameters do you recommend for the LM (beam, alpha, beta)?I found that in the script modeling_psarai.py, lm, alpha=0.5, beta=0.5, beam=16, but in the model card, alpha=0.6, beam_width=32. So I’d like to know what the recommended parameter values are.Thank you.Thank you for releasing the nphearum/PsarAI-S2S model.
In this LM case, I strongly recommended: alpha=0.5, beta=0.5, beam_width=16 (defaults)
| Setting | km | en | fr | mixed | Time |
|---|---|---|---|---|---|
| defaults (0.5 / 0.5 / 16) | 9.26 | 4.64 | 4.33 | 1.57 | 1x |
beam_width=32 |
9.19 | 4.63 | 4.35 | 1.54 | ~1.8x |
alpha=0.6, beam_width=32 |
9.09 | 4.79 | 4.46 | 1.58 | ~1.8x |
beam_width=8 |
9.62 | 4.79 | 4.40 | 1.57 | ~0.5x |