---
license: apache-2.0
language:
- ja
library_name: espnet
tags:
  - automatic-speech-recognition
---

# reazonspeech-espnet-v2

`reazonspeech-espnet-v2` is an automatic speech recognition (ASR) model
trained on [ReazonSpeech v2.0 corpus](https://huggingface.co/datasets/reazon-research/reazonspeech).

## Model Architecture

The general architecture is the same as [reazonspeech-espnet-v1](https://huggingface.co/reazon-research/reazonspeech-espnet-v1).

* Conformer-Transducer model with 118.85M parameters.

* We trained this model for 33 epoch using Adam optimizer. The maximum
  learning rate was 0.02, with 15000 warmup steps.

* The training audio files were sampled at 16khz. Make sure that your
  input audio files have the same sampling rate.

## Usage

We provide `transcribe()` function that is suitable to use with this
model.

```
from espnet2.bin.asr_inference import Speech2Text
from reazonspeech.espnet.asr import transcribe

speech2text = Speech2Text(
    "exp/asr_train_asr_conformer_raw_jp_char/config.yaml",
    "exp/asr_train_asr_conformer_raw_jp_char/valid.acc.ave_10best.pth",
    device="cuda"
)

for cap in transcribe("speech.wav", speech2text):
    print(cap)
```

## License

[Apaceh Licence 2.0](https://choosealicense.com/licenses/apache-2.0/)