Kotoba-Whisper is a collection of distilled Whisper models for Japanese ASR, developed through the collaboration bewteen Asahi Ushio and Kotoba Technologies. Following the original work of distil-whisper (Robust Knowledge Distillation via Large-Scale Pseudo Labelling), we employ OpenAI's Whisper large-v3 as the teacher model, and the student model consists the full encoder of the teacher large-v3 model and the decoder with two layers initialized from the first and last layer of the large-v3 model. Kotoba-Whisper is 6.3x faster than large-v3, while retaining as low error rate as the large-v3.

As successor of our first model, kotoba-whisper-v1.0, we release kotoba-whisper-v2.0 trained on the all subset of ReazonSpeech (the largest speech-transcription paired dataset in Japanese extracted from Japanese TV audio recordings), which amounts 7,203,957 audio clips (5 sec audio with 18 text tokens in average) after those transcriptions more than 10 WER are removed (see WER Filter for detail). The model was trained for 8 epochs with batch size 256 with sampling rate of 16kHz, and the training and evaluation code to reproduce kotoba-whisper is available at https://github.com/kotoba-tech/kotoba-whisper.

Kotoba-whisper-v2.0 achieves better CER and WER than the openai/whisper-large-v3 in the in-domain held-out test set from ReazonSpeech, and achieves competitive CER and WER on the out-of-domain test sets including JSUT basic 5000 and the Japanese subset from CommonVoice 8.0 (see Evaluation for detail).

CER | model | CommonVoice 8 (Japanese test set) | JSUT Basic 5000 | ReazonSpeech (held out test set) | |:--------------------------------------------------------------------------------------------------------------------------------------------------|------------------------------------------------------------------------------------------------------------:|----------------------------------------------------------------------------------------:|------------------------------------------------------------------------------------------------------------:| | kotoba-tech/kotoba-whisper-v2.0 | 9.2 | 8.4 | 11.6 | | kotoba-tech/kotoba-whisper-v1.0 | 9.4 | 8.5 | 12.2 | | openai/whisper-large-v3 | 8.5 | 7.1 | 14.9 | | openai/whisper-large-v2 | 9.7 | 8.2 | 28.1 | | openai/whisper-large | 10 | 8.9 | 34.1 | | openai/whisper-medium | 11.5 | 10 | 33.2 | | openai/whisper-base | 28.6 | 24.9 | 70.4 | | openai/whisper-small | 15.1 | 14.2 | 41.5 | | openai/whisper-tiny | 53.7 | 36.5 | 137.9 |
WER

model	CommonVoice 8 (Japanese test set)	JSUT Basic 5000	ReazonSpeech (held out test set)
kotoba-tech/kotoba-whisper-v2.0	58.8	63.7	55.6
kotoba-tech/kotoba-whisper-v1.0	59.2	64.3	56.4
openai/whisper-large-v3	55.1	59.2	60.2
openai/whisper-large-v2	59.3	63.2	74.1
openai/whisper-large	61.1	66.4	74.9
openai/whisper-medium	63.4	69.5	76
openai/whisper-base	87.2	93	91.8
openai/whisper-small	74.2	81.9	83
openai/whisper-tiny	93.8	97.6	94.9

Latency: As kotoba-whisper uses the same architecture as distil-whisper/distil-large-v3, it inherits the benefit of the improved latency compared to openai/whisper-large-v3 (6.3x faster than large-v3, see the table below taken from distil-whisper/distil-large-v3).

Model	Params / M	Rel. Latency
kotoba-tech/kotoba-whisper-v2.0	756	6.3
kotoba-tech/kotoba-whisper-v1.0	756	6.3
openai/whisper-large-v3	1550	1.0

Transformers Usage

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

bash

pip install --upgrade pip
pip install --upgrade transformers accelerate

Short-Form Transcription

The model can be used with the pipeline class to transcribe short-form audio files (< 30-seconds) as follows:

python

import torch
from transformers import pipeline
from datasets import load_dataset

# config
model_id = "kotoba-tech/kotoba-whisper-v2.0"
torch_dtype = torch.bfloat16 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": "ja", "task": "transcribe"}

# load model
pipe = pipeline(
    "automatic-speech-recognition",
    model=model_id,
    torch_dtype=torch_dtype,
    device=device,
    model_kwargs=model_kwargs
)

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

# run inference
result = pipe(sample, generate_kwargs=generate_kwargs)
print(result["text"])

To transcribe a local audio file, simply pass the path to your audio file when you call the pipeline (make sure the audio is sampled in 16kHz):

diff

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

For segment-level timestamps, pass the argument return_timestamps=True and return the "chunks" output:

python

result = pipe(sample, return_timestamps=True, generate_kwargs=generate_kwargs)
print(result["chunks"])

Sequential Long-Form: Kotoba-whisper is designed to be compatible with OpenAI's sequential long-form transcription algorithm. This algorithm uses a sliding window for buffered inference of long audio files (> 30-seconds), and returns more accurate transcriptions compared to the chunked long-form algorithm. As default, if long audio files are passed to the model, it will transcribes with the sequential long-form transcription. The sequential long-form algorithm should be used in either of the following scenarios:

Transcription accuracy is the most important factor, and latency is less of a consideration
You are transcribing batches of long audio files, in which case the latency of sequential is comparable to chunked, while being up to 0.5% WER more accurate

If you are transcribing single long audio files and latency is the most important factor, you should use the chunked algorithm described below. For a detailed explanation of the different algorithms, refer to Sections 5 of the Distil-Whisper paper. The pipeline class can be used to transcribe long audio files with the sequential algorithm as follows:

Chunked Long-Form

This algorithm should be used when a single large audio file is being transcribed and the fastest possible inference is required. In such circumstances, the chunked algorithm is up to 9x faster than OpenAI's sequential long-form implementation (see Table 7 of the Distil-Whisper paper). To enable chunking, pass the chunk_length_s parameter to the pipeline. For distil-large-v3, a chunk length of 25-seconds is optimal. To activate batching over long audio files, pass the argument batch_size:

python

import torch
from transformers import pipeline
from datasets import load_dataset

# config
model_id = "kotoba-tech/kotoba-whisper-v2.0"
torch_dtype = torch.bfloat16 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": "ja", "task": "transcribe"}

# load model
pipe = pipeline(
    "automatic-speech-recognition",
    model=model_id,
    torch_dtype=torch_dtype,
    device=device,
    model_kwargs=model_kwargs,
    batch_size=16
)

# load sample audio (concatenate instances to create a long audio)
dataset = load_dataset("japanese-asr/ja_asr.reazonspeech_test", split="test")
sample = {"array": np.concatenate([i["array"] for i in dataset[:20]["audio"]]), "sampling_rate": dataset[0]['audio']['sampling_rate']}

# run inference
result = pipe(sample, chunk_length_s=15, generate_kwargs=generate_kwargs)
print(result["text"])

Additional Speed & Memory Improvements

You can apply additional speed and memory improvements to further reduce the inference speed and VRAM requirements. These optimisations primarily target the attention kernel, swapping it from an eager implementation to a more efficient flash attention version.

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:

diff

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

Model Details

See https://huggingface.co/distil-whisper/distil-large-v3#model-details.

Training

Please refer to https://github.com/kotoba-tech/kotoba-whisper for the model training detail. Datasets used in distillation and the whole model variations can be found at https://huggingface.co/japanese-asr.

Evaluation

The following code-snippets demonstrates how to evaluate the kotoba-whisper model on the Japanese subset of the CommonVoice 8.0. First, we need to install the required packages, including 🤗 Datasets to load the audio data, and 🤗 Evaluate to perform the WER calculation:

bash

pip install --upgrade pip
pip install --upgrade transformers datasets[audio] evaluate jiwer

Evaluation can then be run end-to-end with the following example:

python

import torch
from transformers import pipeline
from datasets import load_dataset
from evaluate import load
from transformers.models.whisper.english_normalizer import BasicTextNormalizer

# model config
model_id = "kotoba-tech/kotoba-whisper-v2.0"
torch_dtype = torch.bfloat16 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"}
normalizer = BasicTextNormalizer()

# data config
dataset_name = "japanese-asr/ja_asr.reazonspeech_test"
audio_column = 'audio'
text_column = 'transcription'

# load model
pipe = pipeline(
    "automatic-speech-recognition",
    model=model_id,
    torch_dtype=torch_dtype,
    device=device,
    model_kwargs=model_kwargs,
    batch_size=16
)

# load the dataset and sample the audio with 16kHz
dataset = load_dataset(dataset_name, split="test")
transcriptions = pipe(dataset['audio'])
transcriptions = [normalizer(i['text']).replace(" ", "") for i in transcriptions]
references = [normalizer(i).replace(" ", "") for i in dataset['transcription']]

# compute the CER metric
cer_metric = load("cer")
cer = 100 * cer_metric.compute(predictions=transcriptions, references=references)
print(cer)

The huggingface links to the major Japanese ASR datasets for evaluation are summarized at here. For example, to evaluate the model on JSUT Basic5000, change the dataset_name:

diff

- dataset_name = "japanese-asr/ja_asr.reazonspeech_test"
+ dataset_name = "japanese-asr/ja_asr.jsut_basic5000"

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.