Kotoba-Whisper-Bilingual: kotoba-whisper-bilingual-v1.0 for Whisper cpp

This repository contains the model weights for kotoba-tech/kotoba-whisper-bilingual-v1.0 converted to GGML format. GGML is the weight format expected by C/C++ packages such as Whisper.cpp, for which we provide an example below.

Usage

Kotoba-Whisper can be run with the Whisper.cpp package with the original sequential long-form transcription algorithm.

Steps for getting started:

1. Clone the Whisper.cpp repository:

git clone https://github.com/ggerganov/whisper.cpp.git
cd whisper.cpp
make

2. Download the GGML weights for kotoba-tech/kotoba-whisper-bilingual-v1.0:

wget https://huggingface.co/kotoba-tech/kotoba-whisper-bilingual-v1.0-ggml/resolve/main/ggml-kotoba-whisper-bilingual-v1.0.bin -P ./models

3. Run inference using the provided sample audio:

• Download sample audio. Note that it runs only with 16-bit WAV files, so make sure to convert your input before running the tool. For example, you can use ffmpeg like below.

wget https://huggingface.co/datasets/japanese-asr/en_asr.esb_eval/resolve/main/sample.wav -O sample_en.wav
wget https://huggingface.co/datasets/japanese-asr/ja_asr.jsut_basic5000/resolve/main/sample.flac -O sample_ja.flac
ffmpeg -i sample_en.wav -ar 16000 -ac 1 -c:a pcm_s16le sample_en_fixed.wav
ffmpeg -i sample_ja.flac -ar 16000 -ac 1 -c:a pcm_s16le sample_ja_fixed.wav

• Japanese ASR

make -j && ./main -m models/ggml-kotoba-whisper-bilingual-v1.0.bin -l ja -f sample_ja_fixed.wav --output-file "output.transcribe.ja" --output-json 

• English ASR

make -j && ./main -m models/ggml-kotoba-whisper-bilingual-v1.0.bin -l ja -f sample_en_fixed.wav --output-file "output.transcribe.en" --output-json 

• Japanese (speech) to English (text) Translation

make -j && ./main -m models/ggml-kotoba-whisper-bilingual-v1.0.bin -tr -l en -f sample_ja_fixed.wav --output-file "output.translate.en" --output-json 

• English (speech) to Japanese (text) Translation

make -j && ./main -m models/ggml-kotoba-whisper-bilingual-v1.0.bin -tr -l ja -f sample_en_fixed.wav --output-file "output.translate.ja" --output-json 

Benchmark

We measure the inference speed of different kotoba-whisper-v2.0 implementations with four different Japanese speech audio on MacBook Pro with the following spec:

• Apple M2 Pro
• 32GB
• 14-inch, 2023
• OS Sonoma Version 14.4.1 (23E224)

Scripts to re-run the experiment can be found bellow:

• whisper.cpp
• faster-whisper
• hf pipeline

Currently whisper.cpp and faster-whisper support the sequential long-form decoding, and only Huggingface pipeline supports the chunked long-form decoding, which we empirically found better than the sequnential long-form decoding.

Quantized Model

To use the quantized model, download the quantized GGML weights:

wget https://huggingface.co/kotoba-tech/kotoba-whisper-bilingual-v1.0-ggml/resolve/main/ggml-kotoba-whisper-bilingual-v1.0-q5_0.bin -P ./models

Run inference on the sample audio:

make -j && ./main -m models/ggml-kotoba-whisper-bilingual-v1.0-q5_0.bin -l ja -f sample_ja_fixed.wav --output-file "output.transcribe.ja.q" --output-json 

Note that the benchmark results are almost identical to the raw non-quantized model weight.

Conversion details

The original model was converted with the following command:

# clone OpenAI whisper and whisper.cpp
git clone https://github.com/openai/whisper
git clone https://github.com/ggerganov/whisper.cpp

# get the models
cd whisper.cpp/models
git clone https://huggingface.co/kotoba-tech/kotoba-whisper-bilingual-v1.0

# convert to ggml
python3 ./convert-h5-to-ggml.py ./kotoba-whisper-bilingual-v1.0/ ../../whisper .
mv ggml-model.bin ggml-kotoba-whisper-bilingual-v1.0.bin

# quantize ggml model
cd ../
make quantize
./quantize models/ggml-kotoba-whisper-bilingual-v1.0.bin models/ggml-kotoba-whisper-bilingual-v1.0-q5_0.bin q5_0

Model Details

For more information about the kotoba-whisper-v2.0, refer to the original model card.