Sasayaki

A small CLI tool that simplifies and automates the process of installing and using AI models to transcribe and translate videos. Written in Go, it uses faster-whisper or whisper.cpp for transcription and english translation. Translation into other languages is done using Google Gemini. Just enter the video link or file path to get translated subtitles in .srt format.

The name Sasayaki (ささやき) means "whisper" in Japanese.

Additional Info:

GPU support?

The core features of both implementations download, install and works automatically, and the models run on the CPU by default without requiring any additional system dependencies. Both faster-whisper and whisper.cpp can perform better on GPUs, but this requires manual installation of dependencies and editing the Python script for faster-whisper or manually compiling binaries for whisper.cpp. I plan to add GPU support in the future, but currently, I do not have a dedicated GPU, so I am unable to test and debug GPU functionality. Fortunately, both implementations are fast enough that on a moderately powerful CPU, the medium or small models perform sufficiently well.

Why does this tool exist when there are many similar ones?

Because none of them met my requirements, so I wrote a simple script for myself. Since it works quite well, I decided to refine it slightly and make it available for others. It's just ~700 lines of Go and Python, and I also learned the basics of Golang while creating it. Most moderately advanced users would be able to do it on their own, but I hope it helps someone new or saves someone a bit of time.

Requirements

This tool works on x86_64 Windows and Linux systems, as well as macOS with ARM-based processors, including M1 and newer models. The following dependencies need to be installed and available in your system's PATH:

• ffmpeg
• yt-dlp (optional, only required for downloading videos)

faster-whisper version requires also:

• python
• pyenv

The faster-whisper version does not work on Windows because pyenv is not available for Windows.

Installation

Download version for your operating system from the Releases page. You can rename the file to sasayaki for convenience.

Linux / macOS

Details

Grant the file execute permissions:

chmod +x sasayaki

Then, there are two installation options:

faster-whisper version

Requires python and pyenv.

./sasayaki --install

whisper.cpp version

./sasayaki --install --cpp

both versions at the same time

./sasayaki --install; ./sasayaki --install --cpp

Windows

Details

On Windows faster-whisper version is not available. To install the whisper.cpp version open cmd and run executable:

sasayaki.exe --install

Note

The --install parameter will create a .sasayaki folder in your home directory, next it will install python 3.12 using pyenv, then create a separate venv and download the necessary packages in it. Finally, it will create the python file needed for transcription and a configuration file. If you chose the whisper.cpp version, all Python-related elements will be skipped, and an executable file named whisper-cli will be created in the program directory. You can reverse this process with --uninstall or manually delete the .sasayaki folder, but this will not uninstall a previously installed version of Python from pyenv, you have to do it manually.

The configuration file is located at ~/.sasayaki/config.toml on Linux, macOS and %USERPROFILE%\.sasayaki\config.toml on Windows.

Optional:

• Open config.toml and insert here your Gemini API key
• Set cpu threads and model size in config.toml
• Add sasayaki binary to PATH
• (advanced) Edit transcribe.py to enable running model on GPU (look for commented lines)
• (advanced) Compile whisper.cpp yourself with the parameters that enable GPU acceleration and replace whisper-cli in the program directory with your own executable

Usage

./sasayaki [args] <url>

Possible urls:

• yt-dlp compatibile url
• local video or audio file
• .srt file created by this program (file name must end with " (transcription).srt")

Available args:

  --config
        Use to create or reset config file
  --cpp
        Transcribe using whisper.cpp instead of faster-whisper (enabled by default on Windows)
  --debug
        Print debug info in stdout
  --gemini
        Translate using Google Gemini instead of Whisper
  --install
        Use to install program and needed dependencies in user home folder
  --lang <string>
        Specifies a target translation language when using Google Gemini (default "english")
  --model <string>
        Chose whisper model
  --uninstall
        Use to remove program files and its dependencies from user home folder
  --verbose
        Print commands output in stdout
  --ytdlp
        Download remote video using yt-dlp

Note

The first use of each model (small, medium, large-v3, etc.) requires downloading it. This will happen automatically when the model is selected in the configuration file and transcription is started. The faster-whisper and whisper.cpp models are not compatible with each other and must be downloaded separately. The medium model is approximately 1.4 GiB, while the small model is 461 MiB.

Examples:

# Create english subtitles using only faster-whisper
sasayaki input.mp4

# Create english subtitles using only whisper.cpp
sasayaki --cpp input.mp4

# Create english subtitles using faster-whisper and Google Gemini
sasayaki --gemini input.mp4

# Translate into different language (only with Gemini)
sasayaki --gemini --lang japanese input.mp4

# Download video with yt-dlp then translate it
# The result is a single video file with embedded subtitles.
sasayaki --ytdlp 'example.com/input.mp4'

# There is no need to use --ytdlp for urls starting with "https://" or "http://".
sasayaki 'https://example.com/input.mp4'

# Translate .srt file into another language using Gemini.
# The file name must end with " (transcription).srt"
sasayaki --gemini --lang korean 'input (transcription).srt'

Warning

Each time you use the command with the same video file or link, previously created files will be overwritten.

Compile from source

If you want to use whisper.cpp, you must compile it yourself and then copy the whisper-cli binary into the embed folder.

go mod tidy
go build -ldflags "-w -s"

Changelog

CHANGELOG.md

External Licenses

This project includes a compiled binary file from the whisper.cpp project, which is licensed under the MIT license. The full text of the MIT license can be found in the LICENSE file in the whisper.cpp repository.