benchmark-fastconformer.py

benchmark-fastconformer.py

import torch
import torchaudio

fakewav_80mins = torch.rand(1, 16_000 * 60 * 80)
torchaudio.save("tmp/80mins.wav", fakewav_80mins, 16_000)

import nemo.collections.asr as nemo_asr

asr_model = nemo_asr.models.EncDecCTCModelBPE.from_pretrained(model_name="nvidia/stt_en_fastconformer_ctc_large")
asr_model.eval()

with torch.inference_mode():
    transcriptions = asr_model.transcribe(['tmp/80mins.wav'])

# Change to LongFormer-style local attention
asr_model.change_attention_model(
    self_attention_model="rel_pos_local_attn",
    att_context_size=[128, 128]
)

with torch.inference_mode():
    transcriptions = asr_model.transcribe(['tmp/80mins.wav'])