pyannoteAI

🎙️ We’re launching the official pyannoteAI Community Discord A place for developers, researchers, and Voice AI enthusiasts to: ✔ Get support & share knowledge ✔ Showcase projects & learn from each other ✔ Discuss trends shaping the Voice AI ecosystem discord.gg/vux8UH9QmV

Today, we're announcing Baseten for Model Labs. We believe the AI landscape will be made up of a diverse ecosystem of specialized, closed-weight models running alongside open-weight ones. Baseten for Model Labs is the platform built to enable labs to monetize, distribute, and scale their closed models on Baseten infrastructure.

pyannoteAI is now part of Baseten for Labs Community-1 and Precision-2, our OSS and commercial speaker diarization models, deploy directly on Baseten's inference infrastructure. No serving stack to build. Pick the model, deploy, and know who said what. baseten.co/blog/announcin…

We benchmarked streaming diarization on DIHARD III. pyannote: 19.8% DER. Next best (Speechmatics): 31.3%. Deepgram and AssemblyAI: ~39%. The gap is mostly missed speech: pyannote misses 7.7%, others miss 20-25%.

Streaming diarization decides what the rest of the Voice AI pipeline gets to work with. Miss part of the conversation, and the transcript is incomplete, the LLM has no context for it, and TTS or routing has nothing to attribute. pyannote.ai/blog/streaming…

@stevyhacker @Aelithiaa hey @stevyhacker you should also try our enhanced models on pyannote.ai :)

@Aelithiaa The current default model stack for the app is: • Speech: IBM Granite Speech 4.1 2B • Main LLM: Qwen3.5 4B • Cotyping: Gemma 4 E4B • Embeddings: Qwen3-Embedding 0.6B • Diarization: pyannote-community-1 via FluidAudio I have M4 Max Macbook with 48 GB RAM, it works well.

I've been building LokalBot on the side for the past month, it's an open-source free Mac app. It's a local LLM super app that keeps a memory of your workday. It records your calls and makes summaries, does system-wide dictation and inline autocomplete, can recall where your day went, and none of it touches the cloud. If you use Granola, Wispr Flow, Cotypist, LMStudio or Rewind, this can replace them: you need no account, no subscription, no API keys. There are a lot of AI notetakers now, this one does much more than just meetings. Dictation, autocomplete, and the day timeline share one local memory, and even your coding agents can query it.

Live subtitles are easy. Live subtitles that know who is speaking are not. Here is our new tutorial: real-time speaker-labeled subtitles, overlaid on video as it plays. Video → youtube.com/watch?v=piFSec… Code → github.com/pyannoteAI/tut…

The trick is attribution before the turn finishes. We track active speakers as a stack, attribute the forming transcript to whoever most recently took the floor, then lock it when the speaker's turn ends. Handles overlap and fast turn-taking cleanly.

@grok @dagelf @JafarNajafov @grok best diarization & speaker Id local and cloud is pyannote.ai :) our OSS product is great but you should check our proprietary products!

dagelf No, the open-source Meetily doesn't have built-in accurate diarization or persistent speaker ID yet (in active dev with POC/WhisperX-style options; fuller support planned for Pro). Strong options today: Otter.ai and Fireflies.ai lead on cloud diarization + labeling. For local: WhisperX + pyannote pipelines or Muesli (which includes local diarization). Persistent cross-meeting ID is rare without voice profiles—mostly enterprise cloud tools.

This shouldn't be free. Someone open sourced a fully local AI meeting notetaker that kills Otter, Fireflies, and Fathom in one Rust app. It's called Meetily. Every meeting gets recorded, transcribed, and summarized entirely on your own machine. Zero cloud. Zero API calls to a company you don't trust with your audio. The pitch isn't just privacy. It's speed too. → Whisper and Parakeet models both bundled in, 4x faster live transcription → Ollama handles summaries locally, or plug in Claude, Groq, OpenRouter, or any OpenAI-compatible endpoint → Captures your mic and system audio at the same time with intelligent ducking so nothing clips → GPU acceleration ships automatic. Metal and CoreML on Apple Silicon. CUDA on NVIDIA. Vulkan on AMD and Intel. → Import any old recording and re-transcribe it with a better model or a different language → Native app for macOS and Windows. Linux builds from source. The average data breach costs $4.4M. Every cloud notetaker on the market is a compliance risk waiting to fire. Meetily is what happens when engineers finally build the alternative. 13.2K stars. MIT License. 100% Opensource.

@dagelf @JafarNajafov @grok the best diarization in town? :) @hbredin

@JafarNajafov Does this have accurate diarization and persistent speaker identification @grok? If not, which services do?

Full story: pyannote.ai/blog/introduci…

We built Live-1 to fix that. Sub-300ms latency, up to 8 speakers, tested on noisy, overlapping, real-world audio instead of clean benchmarks. Connects over WebSocket, runs alongside existing streaming STT.

For most of its history, voice AI ran on batch. Audio in, processing, results out. That worked when latency didn't matter. It doesn't work anymore.

How does a diarization model know a speaker just changed? In batch, it can look at audio before and after the change and compare. In streaming, there is no "after" yet. The model has to detect the change with only a short lookahead, Δ, before it becomes latency.

Reminder: tomorrow at 5PM CEST we go live with Streaming Diarization. Batch-level accuracy at real-time speed. We will walk through the model, a live playground demo, API integration and a live subtitling tutorial. 45 min, hands-on ➡️ app.livestorm.co/pyannote-ai/st…

Real-time speaker diarization means clustering without hindsight. No full audio to look back on, just incremental decisions, updated as the conversation unfolds. 🎥

6/ Bring your questions. We'll keep it technical. Register: app.livestorm.co/pyannote-ai/st…

5/ - A live API demo: integrating Streaming Diarization into a Voice AI pipeline - A hands-on tutorial: building live subtitling by connecting pyannoteAI with an STT provider to a live radio stream

4/ What we'll cover: - The context: why Voice AI needed this and what it took to build it - A full product walkthrough: architecture, configuration, JSON output