User Guide

This guide walks through common tasks in the SaLED UI and with the API.

Upload and transcribe audio

1. Open the app at http://localhost:3000.
2. Use the transcription panel to select an audio file (MP3/WAV/M4A).
3. Choose a Whisper model (start with tiny.en for speed, larger models for quality).
4. Enable/disable speaker diarization as needed.
5. Start. Watch progress indicators; segments appear incrementally.

Tips:

• For long files, processing is chunked; progress may advance in bursts.
• If diarization is off, SaLED applies a rule-based fallback to label speakers.

Inspect and navigate transcripts

• Use the waveform and minimap to jump to timestamps.
• The transcript panel shows segments with speaker, start/end, and text.
• Use zoom and playback controls to verify accuracy.

Fix tricky regions with Retranslation

If a region is inaccurate:

1. Note the start_time and end_time of the problematic span.
2. Reprocess just that portion with /api/retranslate/segment by uploading a clipped audio file; or
3. If the original full audio was uploaded for the same job, call /api/retranslate/segment-from-full?job_id=<job_id> with the time range and model.
4. Replace or merge the corrected text in your transcript.

When to escalate model size:

• Background noise, overlapping speakers, or domain-specific jargon.
• Try small.en → medium.en → large-v3(-turbo) as needed.

Summarize a transcript

1. After transcription, send segments to /api/summary/ with an optional full_text.
2. Pick format: paragraph, bullet_points, or key_points.
3. Poll /api/summary/status/{job_id} and fetch /api/summary/{job_id} for the final summary and topics.

Good prompts for summaries:

• Paragraph: broad narrative overview.
• Bullet points: meeting notes or action items.
• Key points: highlights for quick scan.

Chat about the transcript

1. Compose a messages array with chat history and send along transcript/full_text to /api/chat/.
2. Adjust temperature (0.0–1.0) for creativity vs. determinism.
3. Respect token limits with max_tokens for concise answers.

Best practices:

• Ask grounded questions (“What did Speaker 2 request after …?”).
• Provide full_text if possible; it reduces reconstruction overhead.

Managing Whisper models locally (CLI)

From backend/saled:

./cli.py model list
./cli.py model download base.en
./cli.py model delete base.en
./cli.py model update

Keyboard/UX tips

• Use the minimap to scroll large transcripts quickly.
• Toggle diarization for simpler single-speaker content.
• Save versions of transcripts before heavy edits.