reazonspeech-k2-v2 is an automatic speech recognition (ASR) model
trained on
ReazonSpeech v2.0 corpus.
This model provides end-to-end Japanese speech recognition based on
Next-gen Kaldi.
-
Character-based RNN-T model. The total parameter count is 159.34M.
-
This model utilizes an enhanced Transformer architecture called
Zipformer.
-
The training recipe is available on
k2-fsa/icefall.
Note that this model can process Japanese audio clips up to ~30 seconds.
We recommend to use this model through our
reazonspeech
library.
from reazonspeech.k2.asr import load_model, transcribe, audio_from_path
audio = audio_from_path("speech.wav")
model = load_model()
ret = transcribe(model, audio)
print(ret.text)
