whospeaks 0.0.7

Local speaker diarization and realtime voice-labeling tools.

WhoSpeaksLive

Private, speaker-aware transcripts and meeting summaries on hardware you control.

WhoSpeaksLive identifies speakers while a conversation is still happening, produces a stable speaker-labeled transcript, and can turn a saved session into a structured meeting report with a summary, decisions, action items, open questions, risks, and links to supporting transcript evidence.

Unlike meeting bots that depend on participant channels or active-speaker metadata, WhoSpeaksLive identifies speakers from the audio itself - live or recorded, from any source.

Speaker diarization means identifying who spoke when. WhoSpeaksLive combines a fast live path for immediate speaker labels and optional draft text with a separate final path that uses more context for stable sentence-level results.

The core diarization and transcription stack can run entirely on hardware you control. Meeting reports can also use a local or self-hosted large language model (LLM), so sensitive audio, Person-owned Voice samples, transcripts, and summaries can remain inside your environment. The built-in browser server is intended for a trusted operator or protected network; it does not provide authentication or TLS by itself.

Why WhoSpeaksLive

• Meeting intelligence, not just transcription: turn saved speaker-labeled transcripts into evidence-grounded summaries, decisions, action items, questions, and risks.
• Private by design: run locally or on self-hosted servers; no third-party cloud service is required for the core pipeline or local meeting-report generation.
• Useful while it happens: see live speaker labels and optional low-latency draft text while the meeting is still in progress.
• Optional sentence-live translation: translate stable transcript rows into one or several target languages with compact original, single-language, or multi-language views.
• Fast now, stable when final: a low-latency live path gives immediate feedback, while a separate final path uses more context for speaker-labeled transcripts.
• Remember real People safely: link meeting-local Speakers to persistent People, keep a deliberate recognition roster, and confirm suggestions instead of forcing identity.
• Built for review and correction: manage Person-owned Voice samples, correct assignments, and trace report evidence back to the supporting transcript rows.

Use it when the audio is sensitive, the answer is needed immediately, or both: internal meetings, research interviews, legal or medical workflows, local media analysis, and any environment where sending raw conversation audio to another company is not acceptable.

Demo

https://github.com/user-attachments/assets/2de749e0-6c02-47de-b949-bd90b4f4efbb

For faster realtime ASR preview text like shown in the demo, WhoSpeaksLive supports two optional preview backends: Kroko/Banafo streaming models and experimental CPU-only Nemotron 3.5 through sherpa-onnx. Kroko Pro/private models must be installed and licensed separately; Nemotron downloads verified upstream model weights on first use and has separate NVIDIA model-license terms.

Start Here

Install WhoSpeaks and start its native desktop setup and launcher:

pip install whospeaks
whospeaks

The desktop launcher keeps readiness checks, installation, service state, settings, activity, and safe shutdown in one native window. The old interactive terminal launchers are no longer shipped; headless and automated workflows use subcommands such as whospeaks doctor, whospeaks install, whospeaks config, and whospeaks launch.

Or use this same uv command for both the first installation and later updates:

uv tool install --force --python 3.11 whospeaks
uv tool update-shell

Open a new terminal, then run:

whospeaks

--force replaces an older uv installation with the latest stable release. uv tool update-shell only changes PATH when needed, so the two commands are safe to reuse together.

The base package includes PySide6 and the native launcher, while model runtimes and other heavy optional components are installed only after you choose a deployment mode. See PyPI package scope for the exact inclusion and exclusion policy.

The setup application keeps pip as its compatibility default and also offers uv for the larger target-specific install. For automation, pass --installer uv, for example whospeaks install --target local --installer uv --yes. A persistent uv tool install is supported and recommended for a globally available command. Do not use uvx: it creates an ephemeral environment that is not suitable for the runtime packages added by the guided setup.

The desktop launcher opens on Overview. Select the full local stack, the core/controller for remote ASR and embeddings servers, or the ASR/embeddings server packages, then choose whether to include optional realtime preview text. Kroko remains optional because its native runtime may require Python 3.12, Docker Desktop on Windows, or a prebuilt kroko_onnx wheel. Nemotron 3.5 is available as a manual CPU preview option while the installer flow is being integrated.

The whospeaks setup application keeps component readiness, diagnostics, settings, installation progress, logs, cancellation, and browser launch together. See the desktop launcher guide for interface routing, deterministic demo states, and safe-shutdown behavior.

The short whospeaks command is a setup and launcher wrapper. It saves a small profile, runs doctor checks, and expands that profile into the longer whospeaks-window ... browser-server command when you launch.

For a Linux container server:

docker build -t whospeaks:local .
docker run --rm --name whospeaks -p 127.0.0.1:8796:8796 -v whospeaks-data:/data -v whospeaks-models:/models whospeaks:local

See Docker for the full build/run path and volume notes.

For a manual full working setup, follow these in order:

1. Installation: install the Windows controller.
2. External ASR and embeddings servers: set up the Linux GPU services.
3. Quickstart: verify a local or remote smoke run, then run the tuned provider stack.

Current Scope

WhoSpeaksLive performs best on clean recordings where one person speaks at a time into good microphones. Diarization accuracy can degrade with background noise, background music, echo, crosstalk, overlapping speech, or low-quality microphones, and it may become less reliable as the active speaker count grows. The system assumes complete utterances can be assigned to a single speaker, so cases where one speaker starts a sentence and another finishes it are not expected to score well.

CPU-only operation is not the recommended path for the current stack. The system is GPU-heavy today; a CPU-only setup may be possible, but should be treated as a separate optimization target and will likely require engineering work, slower processing, and some accuracy or throughput tradeoffs.

Realtime Preview Languages

Realtime preview text is the optional, low-latency draft transcript shown before final ASR results arrive. Choose one of the following three modes. Set --language or WHOSPEAKS_LANGUAGE to select the language for final ASR and sentence splitting; when preview is enabled, it also selects the preview language.

No realtime preview text

Use --realtime-preview-engine off for the widest language coverage. Live speaker detection and the final speaker-labeled transcript still work; only the rapidly updating draft text is disabled.

Languages: Afrikaans (af), Albanian (sq), Arabic (ar), Armenian (hy), Basque (eu), Belarusian (be), Bulgarian (bg), Catalan (ca), Chinese (zh), Croatian (hr), Czech (cs), Danish (da), Dutch (nl), English (en), Estonian (et), Faroese (fo), Finnish (fi), French (fr), Galician (gl), Georgian (ka), German (de), Greek (el), Hebrew (he or iw), Hindi (hi), Hungarian (hu), Icelandic (is), Indonesian (id), Italian (it), Japanese (ja), Kazakh (kk), Korean (ko), Latin (la), Latvian (lv), Lithuanian (lt), Malayalam (ml), Maltese (mt), Marathi (mr), Myanmar/Burmese (my), Norwegian (no), Norwegian Nynorsk (nn), Persian (fa), Polish (pl), Portuguese (pt), Romanian (ro), Russian (ru), Sanskrit (sa), Serbian (sr), Sindhi (sd), Slovak (sk), Slovenian (sl), Spanish (es), Swedish (sv), Tamil (ta), Telugu (te), Thai (th), Turkish (tr), Ukrainian (uk), Urdu (ur), Vietnamese (vi), and Welsh (cy). These are the languages currently configured for final ASR and at least one sentence segmenter in WhoSpeaksLive.

Kroko

Use --realtime-preview-engine kroko_onnx for Kroko/Banafo streaming preview text.

Languages: German (de), English (en), Spanish (es), French (fr), Italian (it), Hebrew (he or iw), Dutch (nl), Portuguese (pt), Swedish (sv), and Turkish (tr).

Nemotron 3.5

Use --realtime-preview-engine sherpa_onnx for CPU-based Nemotron 3.5 preview text.

Languages: English (en), German (de), Spanish (es), French (fr), Italian (it), Dutch (nl), Portuguese (pt), Turkish (tr), and Swedish (sv). Swedish is available as broad coverage; the other eight are the main supported languages.

Hebrew is not supported by the Nemotron integration; use Kroko or disable realtime preview text. The underlying model may decode more languages, but WhoSpeaksLive treats only the languages above as supported until more have been validated in the realtime path.

See the configuration guide for model selection, downloads, and sentence-segmentation details.

License

WhoSpeaksLive's own code is licensed under the MIT License.

Optional realtime preview support uses separately licensed upstream components and model files. Missing public Kroko Community preview models are downloaded automatically from Hugging Face when Kroko preview starts. Nemotron 3.5 model archives are downloaded from the upstream k2-fsa/sherpa-onnx release and verified with pinned SHA-256 checksums when Nemotron preview starts. This repository's MIT license does not relicense Kroko/Banafo or Nemotron assets; before downloading, bundling, or deploying them, review and respect the current terms from Kroko by Banafo, the Banafo/Kroko-ASR model card, the kroko-ai/kroko-onnx repository, and Third-Party Model Licenses.

Optional translation weights are also downloaded separately and retain their own terms: TranslateGemma uses the Gemma terms, NLLB-200 uses CC-BY-NC-4.0, and MADLAD-400 identifies Apache-2.0. See Live translation and Third-Party Model Licenses.

Documentation

Topic	Document
Documentation map	docs/index.md
Product overview and use cases	docs/overview.md
Installation	docs/installation.md
Desktop setup and launcher	docs/desktop-launcher.md
macOS setup	docs/macos-setup.md
Quickstart	docs/quickstart.md
Live window workflow	docs/live-window-workflow.md
Live translation	docs/translation.md
Meeting intelligence server	docs/meeting-intelligence-server.md
People and cross-meeting recognition	docs/people-and-recognition.md
Security and data privacy	docs/security-and-data-privacy.md
Legacy Speaker-group files	docs/speaker-libraries.md
External ASR and embeddings servers	docs/external-servers.md
Docker server image	docs/docker.md
Configuration guide	docs/configuration.md
Speaker model presets	docs/speaker-model-presets.md
Technical description	docs/technical-description.md
Technical architecture	docs/architecture.md
Validation and scoring	docs/validation-and-scoring.md
Modal deployment	docs/modal-deployment.md
Troubleshooting	docs/troubleshooting.md
Development workflow	docs/development.md
PyPI package scope	docs/pypi-package-scope.md