reazonspeech-nemo-v2 is an automatic speech recognition model trained
on
ReazonSpeech v2.0 corpus.
This model supports inference of long-form Japanese audio clips up to
several hours.
-
Subword-based RNN-T model. The total parameter count is 619M.
-
Encoder uses
Longformer attention
with local context size of 256, and has a single global token.
-
Decoder has a vocabulary space of 3000 tokens constructed by
SentencePiece
unigram tokenizer.
We trained this model for 1 million steps using AdamW optimizer
following Noam annealing schedule.
We recommend to use this model through our
reazonspeech
library.
from reazonspeech.nemo.asr import load_model, transcribe, audio_from_path
audio = audio_from_path("speech.wav")
model = load_model()
ret = transcribe(model, audio)
print(ret.text)
