Kotoba-Whisper-v2.1

Kotoba-Whisper-v2.1 is a Japanese ASR model based on kotoba-tech/kotoba-whisper-v2.0, with additional postprocessing stacks integrated as pipeline. The new features includes (i) improved timestamp achieved by stable-ts and (ii) adding punctuation with punctuators. These libraries are merged into Kotoba-Whisper-v2.1 via pipeline and will be applied seamlessly to the predicted transcription from kotoba-tech/kotoba-whisper-v2.0. The pipeline has been developed through the collaboration between Asahi Ushio and Kotoba Technologies

Following table presents the raw CER (unlike usual CER where the punctuations are removed before computing the metrics, see the evaluation script here) along with the.

model	CommonVoice 8.0 (Japanese)	JSUT Basic 5000	ReazonSpeech Test
kotoba-tech/kotoba-whisper-v2.1 (punctuator + stable-ts)	13.7	11.4	17
kotoba-tech/kotoba-whisper-v2.1 (punctuator)	13.8	11.6	17.3
kotoba-tech/kotoba-whisper-v2.1 (stable-ts)	15.5	15.4	17
kotoba-tech/kotoba-whisper-v2.0	15.4	15.4	17.4
kotoba-tech/kotoba-whisper-v1.1 (punctuator + stable-ts)	13.7	11.2	17.4
kotoba-tech/kotoba-whisper-v1.1 (punctuator)	13.9	11.4	18
kotoba-tech/kotoba-whisper-v1.1 (stable-ts)	15.7	15	17.7
kotoba-tech/kotoba-whisper-v1.0	15.6	15.2	17.8
openai/whisper-large-v3	12.9	13.4	20.6

Regarding to the normalized CER, since those update from v2.1 will be removed by the normalization, kotoba-tech/kotoba-whisper-v2.1 marks the same CER values as kotoba-tech/kotoba-whisper-v2.0.

Latency

Please refer to the section of the latency in the kotoba-whisper-v1.1 here.

Transformers Usage

Kotoba-Whisper-v2.1 is supported in the Hugging Face 🤗 Transformers library from version 4.39 onwards. To run the model, first install the latest version of Transformers.

pip install --upgrade pip
pip install --upgrade transformers accelerate torchaudio
pip install stable-ts==2.16.0
pip install punctuators==0.0.5

Transcription

The model can be used with the pipeline class to transcribe audio files as follows:

import torch
from transformers import pipeline
from datasets import load_dataset

# config
model_id = "kotoba-tech/kotoba-whisper-v2.1"
torch_dtype = torch.float16 if torch.cuda.is_available() else torch.float32
device = "cuda:0" if torch.cuda.is_available() else "cpu"
model_kwargs = {"attn_implementation": "sdpa"} if torch.cuda.is_available() else {}
generate_kwargs = {"language": "japanese", "task": "transcribe"}

# load model
pipe = pipeline(
    model=model_id,
    torch_dtype=torch_dtype,
    device=device,
    model_kwargs=model_kwargs,
    chunk_length_s=15,
    batch_size=16,
    trust_remote_code=True,
    stable_ts=True,
    punctuator=True
)

# load sample audio
dataset = load_dataset("japanese-asr/ja_asr.reazonspeech_test", split="test")
sample = dataset[0]["audio"]

# run inference
result = pipe(sample, return_timestamps=True, generate_kwargs=generate_kwargs)
print(result)

• To transcribe a local audio file, simply pass the path to your audio file when you call the pipeline:

- result = pipe(sample, return_timestamps=True, generate_kwargs=generate_kwargs)
+ result = pipe("audio.mp3", return_timestamps=True, generate_kwargs=generate_kwargs)

• To deactivate stable-ts:

-     stable_ts=True,
+     stable_ts=False,

• To deactivate punctuator:

-     punctuator=True,
+     punctuator=False,

Flash Attention 2

We recommend using Flash-Attention 2 if your GPU allows for it. To do so, you first need to install Flash Attention:

pip install flash-attn --no-build-isolation

Then pass attn_implementation="flash_attention_2" to from_pretrained:

- model_kwargs = {"attn_implementation": "sdpa"} if torch.cuda.is_available() else {}
+ model_kwargs = {"attn_implementation": "flash_attention_2"} if torch.cuda.is_available() else {}

Acknowledgements

• OpenAI for the Whisper model.
• Hugging Face 🤗 Transformers for the model integration.
• Hugging Face 🤗 for the Distil-Whisper codebase.
• Reazon Human Interaction Lab for the ReazonSpeech dataset.