Real-time transcription using faster-whisper
Accepts audio input from a microphone using a Sounddevice. By using Silero VAD(Voice Activity Detection), silent parts are detected and recognized as one voice data. This audio data is converted to text using Faster-Whisper.
The HTML-based GUI allows you to check the transcription results and make detailed settings for the faster-whisper.
If the sentences are well separated, the transcription takes less than a second.
Large-v2 model
Executed with CUDA 11.7 on a NVIDIA GeForce RTX 3060 12GB.
- pip install .
- python -m speech_to_text
- Select "App Settings" and configure the settings.
- Select "Model Settings" and configure the settings.
- Select "Transcribe Settings" and configure the settings.
- Select "VAD Settings" and configure the settings.
- Start Transcription
If you use the OpenAI API for text proofreading, set OPENAI_API_KEY as an environment variable.
- If you select local_model in "Model size or path", the model with the same name in the local folder will be referenced.
- Add generate audio files from input sound.
- Add synchronize audio files with transcription.
Audio and text highlighting are linked.
- Add transcription from audio files.(only wav format)
- Add Send transcription results from a WebSocket server to a WebSocket client.
Example of use: Display subtitles in live streaming.
- Add generate SRT files from transcription result.
- Add support for mp3, ogg, and other audio files.
Depends on Soundfile support. - Add setting to include non-speech data in buffer.
While this will increase memory usage, it will improve transcription accuracy.
- Add non-speech threshold setting.
- Add Text proofreading option via OpenAI API.
Transcription results can be proofread.
- Add feature where audio and word highlighting are synchronized.
if Word Timestamps is true.
- Support for repetition_penalty and no_repeat_ngram_size in transcribe_settings.
- Updating packages.
-
Save and load previous settings.
-
Use Silero VAD
-
Allow local parameters to be set from the GUI.
-
Supports additional options in faster-whisper 0.8.0