About this event
Preparing STT, TTS, and speech-to-speech training data with speaker intelligence
Your voice model is only as clean as its training data
If you train or fine-tune STT, TTS, or S2S models on real conversations, see how speaker intelligence decides what your model learns.
Every voice model starts as hours of podcasts, interviews, and calls that were never recorded to be training data. People interrupt, talk over each other, say "mm-hm" mid-sentence. Before training, all of it has to become clean clips with exactly one voice and boundaries that don't cut words in half.
A 100 ms overlap missed at a turn change is enough to pollute a fine-tuning set, and training can't unlearn it. It shows up later as voice bleed, dropped edge words, and broken turn-taking.
Jyoti Bisht, Developer Relations Engineer at pyannoteAI, will walk through how Voice AI labs build their data pipelines with speaker models such as diarization, speaker identification, and overlap detection, and run it live on a real multi-speaker corpus.
What you'll leave with:
- Where speaker intelligence sits in an 8-step pipeline from raw recordings to training data
- What garbage looks like in audio: missed overlap, cut-off words and speaker confusion, and the effect of each on STT, TTS and S2S models
- Why overlap is noise for TTS but signal for speech-to-speech, and how to prepare data for both
- How to segment by voice before ASR to handle code-switching across languages
- How voiceprints dedupe speakers, cap prolific voices, and keep eval speakers out of training
- How Smallest AI runs 100k+ hours a month through pyannoteAI and removed its post-processing step
- A live demo with Precision-3: from a raw multi-speaker file to training-ready clips
- Answers to your questions in a live Q&A
📅 Date: November 24, 2026, 5 PM CET (11 AM ET, 8 AM PT)
🧑🏫 Format: Talk, live demo, Q&A
⏱️ Duration: about 35 minutes.
🎟️ Register even if you can't attend live. We'll send the recording after the event.
pyannote.ai
Identify who speaks when with pyannote.aiSimply detect, segment, label, and separate speakers in any language.
When your audio sources are complex, contaminated by background noise, affected by external elements, and the performances of your solution depend on them, pyannote brings accuracy and precision.
