parler 🇫🇷

Turn recorded meetings into structured decision logs.

Voxtral transcribes. Mistral extracts. parler delivers decisions, commitments, and reports — in French and beyond.

Studio — live pipeline run, two-model stage strip, real-time progress	Results — structured decision log with decisions, commitments, and run metadata
Artifacts — project file tree and local cache entry browser	About — capability table, pipeline stage legend, and keyboard shortcuts

flowchart LR
    A["🎙️ Audio\nmp3 · wav · m4a\nflac · ogg · aac"]

    subgraph pipeline ["  parler pipeline  "]
        direction LR
        B["Ingest\nprobe & normalize"]
        C["Transcribe\n✦ Voxtral"]
        D["Attribute\nspeaker labels"]
        E["Extract\n✦ Mistral"]
        F["Render\nMarkdown · HTML · JSON"]
    end

    G["📄 Decision Log\ndecisions · commitments\nquestions · rejected"]

    A --> B --> C --> D --> E --> F --> G

Two AI models (by @mistralai) power the pipeline:

Every run writes a local checkpoint. Re-render or re-extract at any time, no API call required.

1. Install

git clone https://github.com/AbdelStark/parler && cd parler
uv sync --locked --group dev
cp .env.example .env          # add MISTRAL_API_KEY=sk-...

2. Validate

uv run parler doctor

✓ API key      MISTRAL_API_KEY found
✓ Config       parler.toml readable
✓ Artifacts    .parler-cache · .parler-runs writable
✓ FFmpeg       ffprobe available

3. Run

uv run parler process tests/fixtures/audio/fr_meeting_5min.mp3 \
  --lang fr \
  --participant Pierre \
  --participant Sophie \
  --meeting-date 2026-04-09

fr_meeting_5min-decisions.md

uv run parler process meeting.mp3 \
  --lang fr,en \
  --participant Pierre \
  --participant Alice \
  --output decisions.md

• --lang — expected language codes; auto-detected if omitted
• --participant — known speaker name; repeat for each person
• --output — destination path; extension sets the format (.md / .html / .json)
• --local — run transcription and extraction with the local Hugging Face mistralai/Voxtral-Mini-3B-2507 model instead of the hosted Mistral API
• --verbose — emit stage-by-stage progress, model names, trace ID, and output summaries to stderr while keeping stdout reserved for the final output path

uv run parler transcribe meeting.mp3 --format json --output transcript.json

Install the local inference stack first:

uv add 'torch>=2.4' 'transformers>=4.57' 'mistral-common[audio]'

Then run the pipeline with --local:

uv run parler process meeting.mp3 \
  --lang fr \
  --participant Pierre \
  --participant Sophie \
  --meeting-date 2026-04-09 \
  --local \
  --verbose

Verbose local mode shows the pipeline stages, selected models, and run metadata as they execute.	The generated markdown report shows the extracted decisions, commitments, questions, and rejections offline.

parler maps --local to the open-weights mistralai/Voxtral-Mini-3B-2507 model for both transcription and transcript-to-JSON extraction. The Hugging Face mistralai/Voxtral-4B-TTS-2603 repo is a text-to-speech model, not the speech-to-text model used here. Local transcription also depends on the mistral-common[audio] stack that Hugging Face's Voxtral processor uses to build transcription prompts.

uv run parler report --from-state .parler-state.json --format html --output report.html

uv run parler tui

Boots with a synthetic French demo preloaded. Press Ctrl+R to run the full pipeline immediately, or pick a real VoxPopuli FR clip from the sidebar.

Some of the sample audio files used by parler were generated from clips in the Meta AI Research VoxPopuli dataset.

# Decision Log

Meeting: 2026-04-09  ·  Languages: fr  ·  Generated by: parler · mistral-medium-latest · v0.1.0

## Decisions (1)

| ID | Summary                                             | Owner  | Timestamp | Confidence |
|----|-----------------------------------------------------|--------|-----------|------------|
| D1 | On part sur le 15 mai pour le lancement. C'est décidé. | Pierre | 03:42 | high |

## Commitments (2)

| ID | Owner  | Action                          | Deadline   | Confidence |
|----|--------|---------------------------------|------------|------------|
| C1 | Sophie | Préparer la démo technique      | 2026-04-12 | high       |
| C2 | Pierre | Envoyer l'invite à l'équipe     | 2026-04-10 | medium     |

[transcription]
model    = "voxtral-mini-latest"    # voxtral-small-latest for higher accuracy
languages = ["fr"]

[extraction]
model = "mistral-medium-latest"     # mistral-large-latest for harder meetings

[cache]
directory = ".parler-cache"

[cost]
max_usd = 2.00                      # abort if preflight estimate exceeds this

# Estimate spend before any API call
uv run parler process meeting.mp3 --cost-estimate

# Inspect cache
uv run parler cache list
uv run parler cache show <key>
uv run parler cache clear --yes

# Inspect run traces
uv run parler runs list
uv run parler runs show <trace_id>

# Prune stale artifacts
uv run parler cleanup --older-than-days 7

parler ships tab-completion snippets for bash, zsh, and fish. Pipe the relevant snippet into your shell config:

# bash
parler completion bash >> ~/.bash_completion

# zsh
parler completion zsh >> ~/.zshrc

# fish
parler completion fish > ~/.config/fish/completions/parler.fish

After reloading your shell, tab-completion works for all parler subcommands and options.

# Fast verification slice
uv run pytest tests/unit tests/integration tests/property -q

# Full quality gate
uv run ruff check parler
uv run mypy parler/
uv build

MIT