voicelog

The changelog that talks to you. At any git moment — after a git pull, before opening a PR, or just to catch up — voicelog gives you a short spoken + written rundown of what changed. Both the LLM and the voice are configurable (NVIDIA, OpenAI, ElevenLabs, and more). Run voicelog, and it prints a casual changelog and reads a summary aloud.

Install

pip install voicelog            # core (text rundown; also all you need for OpenAI/ElevenLabs speech)
pip install "voicelog[tts]"     # + NVIDIA Riva TTS support (only needed for the riva provider)

Requires Python 3.10+. Speech playback works on Windows, macOS, and Linux (via winsound, afplay, and paplay/aplay/ffplay respectively). The [tts] extra pulls in NVIDIA's gRPC client — only needed if you use tts_provider: riva (the default). If you set tts_provider: openai or elevenlabs, plain pip install voicelog is enough since those just use HTTP.

Verify it's installed:

voicelog --help

Install from source (for development)

git clone https://github.com/Navaneethp007/voicelog.git
cd voicelog
pip install -e ".[tts,dev]"
pytest

Setup — one API key

voicelog needs a free API key from build.nvidia.com (no credit card). You can either let voicelog ask you on first run, or set it yourself.

Easiest: just run voicelog in a repo — if no key is set, it prompts you to paste one and uses it for that session.

Or set it manually (the same key powers text and speech):

export NVIDIA_API_KEY=nvapi-...     # bash / macOS / Linux
set NVIDIA_API_KEY=nvapi-...        # Windows cmd (this session)
$env:NVIDIA_API_KEY = "nvapi-..."   # PowerShell (this session)

To persist it on Windows so every terminal has it: setx NVIDIA_API_KEY "nvapi-..." (then open a new terminal).

Usage

# From inside your git repo: prints the changelog AND speaks it
voicelog

# Print only, no audio
voicelog --no-speak

# Ignore the cache and regenerate fresh prose from the model
voicelog --fresh

# Use an alternate config file
voicelog --config path/to/changelog.yml

Catch up on a git pull

Just pulled a branch and want to know what landed? --pull summarizes exactly the commits the pull brought in (everything since ORIG_HEAD) — text + voice:

git pull
voicelog --pull          # "here's what you just pulled", spoken + printed

Preview a PR before you open it

On a feature branch, --pr summarizes what your branch would put in a pull request — the commits since the base branch. The base is auto-detected (origin/HEAD, falling back to main/master), or you can name it:

voicelog --pr           # against the auto-detected base
voicelog --pr develop   # against a base branch you choose

Diff against any ref

Or compare against any ref with --since:

voicelog --since main              # commits since main
voicelog --since v1.2.0            # commits since a tag
voicelog --since HEAD~10           # last 10 commits

These are transient "what changed" views — they print and speak, but do not touch the persistent .changelog/voice.md (that's reserved for releases).

Orient yourself on a repo you just cloned

--new gives you a spoken + written overview of a project you're new to: what it is (from the README) plus what's been happening lately (the most recent commits, regardless of tags/releases):

voicelog --new

Reads onboard_commits recent commits (default 15 — tune it in changelog.yml) and the repo's README (README.md/.rst/.txt). Missing a README? It still orients you from the commits alone. Like the other modes above, --new is transient — it never writes .changelog/voice.md.

Richer summaries with --with-diff

By default voicelog sends only commit messages, changed filenames, and a diffstat (lines added/removed) to the model — never your actual code. If commit messages are vague and you want the model to understand what really changed, opt in to sending the real code diff:

voicelog --pull --with-diff

This prints a privacy warning every time, since it's the one thing in voicelog that sends code to the LLM provider. Works with every mode (default, --pull, --pr, --since, --new).

More (or less) spoken detail

By default the spoken summary is brief (2-3 sentences). For a fuller walkthrough of each change, add --detail:

voicelog --pull --detail

Set speech_detail: detailed in changelog.yml to make it the default. (The printed text is always the full changelog — this only changes the spoken part.)

Caching

The generated changelog is cached on the set of commits (in .changelog/.voicelog-cache.json). Re-running with the same commits returns the cached text instead of calling the model again — so repeat runs are free, fast, and don't churn voice.md. Add a new commit (or pass --fresh) to regenerate. You'll usually want to git-ignore the cache file.

By default, each run is a transient rundown — it prints the casual changelog and reads a short spoken summary aloud, but doesn't write anything. That's the everyday use: quick "what changed" at a git moment.

Each run:

1. Prints the full casual changelog (under ## Unreleased) to stdout.
2. Reads a short spoken summary aloud via NVIDIA Riva TTS (2-3 sentences — not the whole changelog, so audio stays quick even on huge repos).

Add --changelog (or set write_changelog: true) to also maintain a persistent release changelog at .changelog/voice.md that accumulates across releases (with automatic promotion when you tag). That's opt-in — most runs don't need it.

Working on large repos

For big histories (e.g. a fork with hundreds of commits since the last tag), voicelog caps how many commits it sends to the model — the most recent max_commits (default 50) — so generation stays fast and cheap. The spoken part is always a short summary, so it never gets stuck reading a giant changelog aloud. Tune max_commits in changelog.yml, or use --no-speak for instant text-only output.

Persistent changelog (.changelog/voice.md)

voicelog maintains this file for you:

• Keeps the current ## Unreleased block at the top, refreshed each run (idempotent — running twice with no new commits changes nothing).
• When you git tag a release, the previous ## Unreleased is promoted to ## <tag> and a fresh ## Unreleased starts above it. Your history accumulates; nothing is lost.

Using your own LLM

voicelog talks to any OpenAI-compatible endpoint. Set three things in changelog.yml — base_url, model, and api_key_env (the name of the env var your key lives in) — then export that key. The key is never stored in the file.

Provider	base_url	model	api_key_env
NVIDIA (default)	https://integrate.api.nvidia.com/v1	mistralai/mistral-medium-3.5-128b	NVIDIA_API_KEY
OpenAI	https://api.openai.com/v1	gpt-4o-mini	OPENAI_API_KEY
Groq (fast, free tier)	https://api.groq.com/openai/v1	llama-3.1-8b-instant	GROQ_API_KEY
Ollama (local, no key)	http://localhost:11434/v1	llama3.1	any name

Using your own voice

The spoken part is just as configurable — set tts_provider in changelog.yml. All three work cross-platform (Windows/macOS/Linux) and are independent of your text LLM choice above (mix and match freely, e.g. Groq for text + ElevenLabs for voice).

Provider	tts_provider	tts_voice	tts_api_key_env
NVIDIA Riva (default)	riva	Magpie-Multilingual.EN-US.Sofia	NVIDIA_API_KEY
OpenAI TTS	openai	alloy (or echo/fable/onyx/nova/shimmer)	OPENAI_API_KEY
ElevenLabs	elevenlabs	your ElevenLabs voice id	ELEVENLABS_API_KEY

No key for any of them? Set speak: false — the text changelog still works everywhere. See the comments in changelog.yml for the full field list per provider (tts_model, tts_base_url, etc).

Configuration

Copy changelog.yml and edit it — all fields are optional, built-in defaults work out of the box. Key fields:

Field	Purpose
model	Any model on build.nvidia.com (default mistral-medium)
sections	Section grouping for the changelog
noise	Regex commit subjects to drop (e.g. ^wip)
speak	true/false — read aloud (also --no-speak per run)
tts_provider	riva / openai / elevenlabs
tts_voice	Voice id/name — meaning depends on the provider
voice_md	Path to the persistent changelog
max_commits	Cap on commits sent to the model for normal ranges
onboard_commits	Recent commits --new reads (ignores tags)

Every secret is read from an environment variable named in the config (api_key_env for text, tts_api_key_env for speech) — never stored in the file itself.

Graceful degradation

voicelog never blocks on audio. Text always prints first; if anything downstream fails it only warns to stderr:

• LLM unavailable → prints a plain bulleted commit list instead.
• TTS misconfigured / key missing / no audio player / synthesis error → prints text, warns, skips speech.

Architecture

Pure data pipeline:

gitsource → filters → voice → prompt → llm → render → cli ─┬─ stdout
  (or readme +                                            ├─ voicefile (.changelog/voice.md, opt-in)
   read_recent_commits                                     └─ tts (speak aloud)
   for --new)

llm and tts are the only provider-aware modules. Swapping provider/model/voice is a config change, never a code change.

Tests

pytest