Sonix

Promptfree, for everyone, and the only version there is

Copy the prompt for Claude CodeCopy the prompt for CodexCopy the prompt for Cursor

Copy

Build me the transcription service that replaces Sonix.

This moved from expensive to free: local models now transcribe well on an ordinary machine. **The cost is time, not money**, and the editor is where the remaining work is.

STACK
- Node 20+ with Fastify, server-rendered HTML with a little vanilla JS
- SQLite through better-sqlite3, WAL mode, FTS5 over transcripts
- whisper.cpp for transcription, ffmpeg for extraction and conversion
- Caddy in front

THE DATA MODEL
- media: id, path, sha256, duration_ms, language, status, uploaded_at
- segments: id, media_id, index, speaker, start_ms, end_ms, text, edited_text, confidence
- speakers: id, media_id, label, name
- exports: id, media_id, format, path, created_at
- **Word-level timings stored, not just segment timings.** Everything good in the editor depends on them, and adding them later means re-transcribing everything

THE PIPELINE
- ffmpeg to 16kHz mono WAV, which is what the model wants and what makes it fast
- whisper.cpp with a model chosen by need: the medium model is a good default, the large one is better and several times slower
- A queue with a concurrency cap; transcription saturates every core it is given
- Long files split on silence, transcribed in parallel, and stitched with the offsets corrected
- Detect the language, and let it be overridden — detection is wrong often enough to matter on short or accented audio

THE EDITOR, WHICH IS THE REAL PRODUCT
- Text on the left, media playing along, the current word highlighted. Click a word, the audio jumps there
- Edits stored against the segment with the original kept, so the machine output and the human correction are both on record
- **A shortcut to play, pause and rewind three seconds without leaving the keyboard.** That single shortcut is most of the difference between fast correction and slow
- Low-confidence words marked, so attention goes where it is needed
- Find and replace across the transcript, with a per-speaker option

SPEAKERS
- Diarise into unlabelled speakers, then name them once and apply everywhere
- Merge and split speakers, because diarisation always over-splits somewhere

EXPORT, WHICH MUST BE COMPLETE
- SRT and VTT with a proper reading-speed limit and line length — **subtitles that flash for half a second are unreadable**, so cap characters per line and enforce a minimum duration
- Plain text with and without timestamps, and a Word-friendly format with speakers
- Burnt-in subtitles through ffmpeg as an option

TRANSLATION
- Translate segment by segment, keeping the timings, and mark translated text as machine output
- Never overwrite the source language transcript

WHAT MATTERS MOST
Word-level timings and the keyboard shortcuts. Together they make correction fast enough that the whole thing is worth doing yourself.