Why local speech-to-text matters
Cloud speech APIs are convenient until privacy, cost, or offline policy gets in the way. Interviews, meeting notes, podcast drafts, and lecture recordings often should not leave your infrastructure. Fah Swe Tools offers Speech to Text powered by local Whisper via faster-whisper when the server has it installed — not an OpenAI cloud call from the browser.
If the model or package is missing, you get a clear install error instead of a fake empty transcript.
How to transcribe audio or video
- Open Speech to Text.
- Upload MP3, WAV, M4A, MP4, WebM, or similar (media upload limits apply).
- Pick a language or leave Auto-detect (English, Bangla, and many others Whisper supports).
- Run transcription.
- Copy the text or download the
.txt. Files auto-delete within 24 hours.
Server requirements (for operators)
On the host (e.g. via PuTTY):
pip install faster-whisper
# ffmpeg must be on PATH for many containers
# First run downloads WHISPER_MODEL (default: tiny)
export WHISPER_MODEL=tiny # or base / small if you have RAM
- ffmpeg — recommended for extracting audio from video
- Disk/network — first model download (~75 MB for
tiny) - CPU — int8 compute; keep
tinyorbaseon shared VPS hardware
Quality tips for better transcripts
- Prefer clear speech, low music bed, and a single primary speaker
- Trim long silences with Trim Audio first
- Extract a clean track with Extract Audio when the video is huge
- For images of documents, use OCR instead
Privacy stance
Uploads are processed and typically deleted within 24 hours. Transcription runs on the tools server’s Whisper stack when installed — we do not market this as a third-party SaaS speech API.
FAQ
Does this work without faster-whisper?
No. Installfaster-whisper (and preferably ffmpeg). The UI shows a clear error if missing.
Bangla support?
Whisper can auto-detect Bangla; pickbn when you know the language for more stable results.