←
pyannoteAI152 posts
pyannoteAI avatar

pyannoteAI

@pyannoteAI

Make the most of conversational speech with state of the art voice AI

◷ Joined April 2024▧ 34 Photos and videos
pyannoteAI avatar
pyannoteAI@pyannoteAI·9 Oct 2025

🎙️ 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

Post attachment
pyannoteAI avatar
pyannoteAI@pyannoteAI·1h

pyannote.audio started as a research toolkit for speaker diarization. Today, more than 400,000 people have downloaded pyannote models from Hugging Face. Thank you to every contributor and everyone who opened an issue.

Post attachment
pyannoteAI avatar
pyannoteAI@pyannoteAI·Nov 12

At pyannoteAI → We never train on your audio, your outputs or any customer data. On every plan, with nothing to opt out of. → Need it to stay in Europe? Switch your processing region to EU. Your code doesn't change. → Need it to never leave your network? Run us on-premise

pyannoteAI avatar
pyannoteAI@pyannoteAI·Nov 10

VapiCon 2026 starts tomorrow 🚀 pyannoteAI booth at Fort Mason, Nov 11-12. Hervé on stage Thursday, 1:40 PM. Building voice apps and care about accuracy on real-world audio? Come by.

Post attachment
pyannoteAI avatar
pyannoteAI@pyannoteAI·Nov 9

TTS and speech-to-speech need opposite things from the same audio. For TTS, overlap is noise to cut out. For S2S, it's the signal that teaches turn-taking. Same speaker layer, two ways to use it.

pyannoteAI avatar
pyannoteAI@pyannoteAI·Nov 9

Three failure modes we see in training pipelines: Missed overlap → STT hallucinates merged words; S2S learns broken turn-taking. Cut-off words → TTS gets clipped onsets. Speaker confusion → TTS learns an averaged, unstable voice

pyannoteAI avatar
pyannoteAI@pyannoteAI·Nov 9

A 100 ms overlap at a turn change. Missed by segmentation, labeled as single-speaker, repeated across thousands of clips. That's how voice bleed and dropped words end up in a model trained on "clean" data. ➡️ Webinar on Voice AI training data, Nov 24, 5 PM CET. Thread below.

pyannoteAI avatar
pyannoteAI@pyannoteAI·Nov 6

pyannote models work with any STT, so the program fits alongside the speech-to-text provider you already use. To apply, you should have fewer than 25 people, a product live or launching within 6 months, and voice at the core of what you build.

Post attachment
pyannoteAI avatar
pyannoteAI@pyannoteAI·Nov 6

What accepted startups get: • Up to $3,000 in API credits, valid for 12 months • An onboarding session with a solution engineer on your pipeline • Early access to new models before public release • Co-marketing through customer stories and joint technical content