Kotoba-Whisper: kotoba-whisper-v2.0 for Whisper cpp

This repository contains the model weights for kotoba-tech/kotoba-whisper-v2.0 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:

Clone the Whisper.cpp repository:

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

Download the GGML weights for kotoba-tech/kotoba-whisper-v2.0:

wget https://huggingface.co/kotoba-tech/kotoba-whisper-v2.0-ggml/resolve/main/ggml-kotoba-whisper-v2.0.bin -P ./models

Run inference using the provided sample audio:

wget https://huggingface.co/kotoba-tech/kotoba-whisper-v1.0-ggml/resolve/main/sample_ja_speech.wav
make -j && ./main -m models/ggml-kotoba-whisper-v2.0.bin -l ja -f sample_ja_speech.wav --output-file transcription --output-json

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 this:

ffmpeg -i input.mp3 -ar 16000 -ac 1 -c:a pcm_s16le output.wav

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)

audio file	audio duration (min)	whisper.cpp (sec)	faster-whisper (sec)	hf pipeline (sec)
audio 1	50.3	581	2601	807
audio 2	5.6	41	73	61
audio 3	4.9	30	141	54
audio 4	5.6	35	126	69

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

whisper.cpp
faster-whisper
hf pipeline
Also, 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-v2.0-ggml/resolve/main/ggml-kotoba-whisper-v2.0-q5_0.bin -P ./models

Run inference on the sample audio:

make -j && ./main -m models/ggml-kotoba-whisper-v2.0-q5_0.bin -l ja -f sample_ja_speech.wav --output-file transcription.quantized --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-v2.0

# convert to ggml
python3 ./convert-h5-to-ggml.py ./kotoba-whisper-v2.0/ ../../whisper .
mv ggml-model.bin ggml-kotoba-whisper-v2.0

# quantize ggml model
cd ../
make quantize
./quantize models/ggml-kotoba-whisper-v2.0.bin models/ggml-kotoba-whisper-v2.0-q5_0.bin q5_0

Model Details

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

kotoba-tech
/

kotoba-whisper-v2.0-ggml