How to suppress tokens for a EncDecHybridRNNTCTCBPEModel model?

[kryvokhyzha]

Hi!

I'm using the following NeMo ASR multilingual model:

import nemo.collections.asr as nemo_asr
import copy

AUDIO_FILE_PATH = "sample.wav"

asr_model = nemo_asr.models.EncDecHybridRNNTCTCBPEModel.from_pretrained(
    model_name="stt_multilingual_fastconformer_hybrid_large_pc",
)

decoding_cfg = copy.deepcopy(asr_model.cfg.decoding) 
decoding_cfg.strategy = "beam"
decoding_cfg.beam.beam_size = 25
decoding_cfg.beam.return_best_hypothesis = True

asr_model.change_decoding_strategy(decoding_cfg)

print(asr_model.transcribe([AUDIO_FILE_PATH]))

What do I want?

I want to suppress a list of tokens for all languages except Ukrainian and English. Is there a way to achieve this?

Expected Behavior

• Prevent the model from predicting tokens belonging to other languages.
• Ensure that only Ukrainian and English tokens are recognized.

Analogues

• suppress_tokens argument in faster-whisper