Skip to content

Latest commit

 

History

History
135 lines (101 loc) · 3.74 KB

transcribe.md

File metadata and controls

135 lines (101 loc) · 3.74 KB

Transcribe — Offline Speech-to-Text

Convert recordings to text using a built-in Whisper engine. Fully offline — no external binaries, no cloud APIs.

Quick start

plaude transcribe recording.wav
# First run auto-downloads the model (~369 MB)

Quality presets

plaude transcribe recording.wav                  # default: medium
plaude transcribe --quality fast recording.wav    # fastest, ~75 MB model
plaude transcribe --quality high recording.wav    # best accuracy, ~1.5 GB model
Preset Speed Accuracy Model size Best for
fast ~2s / 30s audio Good 75 MB Quick notes, testing
medium ~5s / 30s audio Very good 369 MB Daily voice memos
high ~15s / 30s audio Best 1.5 GB Important meetings

Speaker identification

Identify who said what:

plaude transcribe --diarize recording.wav
# [Speaker 1] [00:00:00.00 → 00:00:12.40] Hello, this is a test.
# [Speaker 2] [00:00:12.40 → 00:00:25.20] Thanks for the demo.

Diarization models (~53 MB) are downloaded automatically on first use.

Output formats

plaude transcribe recording.wav                          # plain text (default)
plaude transcribe --output-format srt recording.wav      # SubRip subtitles
plaude transcribe --output-format vtt recording.wav      # WebVTT subtitles
plaude transcribe --output-format json recording.wav     # structured JSON

JSON output (for AI pipelines)

plaude transcribe --output-format json --diarize recording.wav
{
  "file": "recording.wav",
  "duration_seconds": 25.2,
  "language": "auto",
  "quality": "medium",
  "model": "ggml-distil-medium.en.bin",
  "segments": [
    {"start": 0.0, "end": 12.4, "text": "Hello...", "speaker": "Speaker 1"},
    {"start": 12.4, "end": 25.2, "text": "Thanks...", "speaker": "Speaker 2"}
  ],
  "speakers": ["Speaker 1", "Speaker 2"],
  "full_text": "Hello... Thanks..."
}

Piping to AI tools:

plaude transcribe --output-format json recording.wav | jq -r '.full_text' | llm "Summarize this"

Language support

Default presets use English-only models. For other languages:

plaude transcribe --language de recording.wav

All options

plaude transcribe [OPTIONS] <FILES>...

Arguments:
  <FILES>...              WAV files to transcribe

Options:
  --quality <PRESET>      fast, medium (default), high
  --output-format <FMT>   txt (default), srt, vtt, json
  --language <CODE>       Language hint (e.g. en, de, ja)
  --diarize               Identify speakers
  --model <PATH>          Custom GGML model file
  --models-dir <PATH>     Model cache directory
  --list-models           Show available models
  --no-download           Don't auto-download missing models

Model management

Models are cached at ~/.local/share/plaude/models/. Override with PLAUDE_MODELS_DIR.

plaude transcribe --list-models    # show all available models with download URLs

Full workflow

plaude record start
# ... speak ...
plaude record stop
plaude files list
plaude files pull-one <ID> -o ~/plaud
plaude transcribe --quality high ~/plaud/<ID>.wav > ~/plaud/<ID>.txt

# Summarize the transcript
plaude summarize ~/plaud/<ID>.txt
plaude summarize --template action-items ~/plaud/<ID>.txt

# Fix transcription errors (optional)
plaude correct ~/plaud/<ID>.txt

See LLM Integration for full summarization and correction documentation.

Progress and tips

  • Phase indicators: Loading model → Transcribing → Identifying speakers (if --diarize)
  • Auto-tip: If transcription takes >30s, suggests --quality fast
  • JSON mode: Progress suppressed on stderr — stdout stays clean for piping
  • Multiple files: Per-file progress with filename prefix