This is a bash script that utilizes the OpenAI Whisper API to transcribe continuous voice input into text. It uses SoX for audio recording and includes a built-in feature that detects silence between speech segments.
The script is designed to convert voice audio into text each time the system identifies a specified duration of silence. This enables the Whisper API to function as if it were capable of real-time speech-to-text conversion. It is also possible to specify the audio file to be converted by Whisper.
After transcription, the text is automatically copied to your system's clipboard for immediate use. It can also be saved in a specified directory as a text file.
Run the following commands:
> brew tap yohasebe/whisper-stream
> brew install whisper-stream
That's it! The whisper-stream command should now be available in your terminal.
- Install the following dependencies:
curljqsoxxclip(for Linux)alsa-utils(optional for Linux)
On a Debian-based Linux distribution, you can install these dependencies with:
sudo apt-get install curl jq sox xclip alsa-utils- Identify a directory in your system's PATH variable where you want to place the script. You can check the directories in your PATH variable by running the following command:
> echo $PATH- Move the
whisper-streamscript to the chosen directory. For example, if you want to move it to/usr/local/bin, run the following command:
> mv whisper-stream /usr/local/bin- Make sure the script is executable by running the following command:
> chmod +x /usr/local/bin/whisper-streamYou can start the script with the following command:
> whisper-stream [options]The available options are:
-v, --volume <value>: Set the minimum volume threshold (default: 1%)-s, --silence <value>: Set the minimum silence length (default: 1.5)-o, --oneshot: Enable one-shot mode-d, --duration <value>: Set the recording duration in seconds (default: 0, continuous)-t, --token <value>: Set the OpenAI API token-p, --path <value>: Set the output directory path to create the transcription file-g, --granularities <value>: Set the timestamp granularities (segment or word)-r, --prompt <value>: Set the prompt for the API call-l, --language <value>: Set the input language in ISO-639-1 format-f, --file <value>: Set the audio file to be transcribed-tr, --translate: Translate the transcribed text to English-p2, --pipe-to <cmd>: Pipe the transcribed text to the specified command (e.g., 'wc -w')-q, --quiet: Suppress the banner and settings-V, --version: Show the version number-h, --help: Display the help message
Here are some usage examples with a brief comment on each of them:
> whisper-stream
This will start the script with the default settings, recording audio continuously and transcribing it into text using the default volume threshold and silence length. If the OpenAI API token is not provided as an argument, the script will automatically use the value of the OPENAI_API_KEY environment variable if it is set.
> whisper-stream -l ja
This will start the script with the input language specified as Japanese; see the Wikipedia page for ISO-639-1 language codes.
> whisper-stream -tr
It transcribes the spoken audio in whatever language and presents the text translated into English. Currently, the target language for translation is limited to English.
> whisper-stream -v 2% -s 2 -o -d 60 -t your_openai_api_token
This example sets the minimum volume threshold to 2%, the minimum silence length to 2 seconds, enables one-shot mode, sets the recording duration to 60 seconds, and specifies the OpenAI API token.
> whisper-stream -f ~/Desktop/interview.mp3 -p ~/Desktop/transcripts -l en
This will transcribe the audio file located at ~/Desktop/interview.mp3. The input language is specified as English. The output directory is set to ~Desktop/transcripts to create a transcription text file.
> whisper-stream -p2 'wc -w'
This will start the script with the default settings for recording audio and transcribing it. After transcription, the transcribed text will be piped to the wc -w command, which counts the number of words in the text. The result, indicating the total word count, will be printed below the original transcribed output.
> whisper-stream -v segment -p ~/Desktop
The -g option allows you to specify the mode for timestamp granularities. The available modes are segment or word, and specifying either will display detailed transcript data in JSON format. When used in conjunction with the -p option to specify a directory, the results will be saved as a JSON file. For more information, see the timestamp_granularities[] section in OpenAI Whisper API reference.
Restrictions such as the languages that can be converted by this program, the types of audio files that can be input, and the size of data that can be converted at one time depend on what the Whisper API specifies. Please refer to Whisper API FAQ.
Yoichiro Hasebe [yohasebe@gmail.com]
This software is distributed under the MIT License.