Meta's AI Separates 20+ Speakers in Real-Time Audio

Why this matters

For developers building call-center analytics, meeting tools, or multilingual assistants, the math here is simple: $0.18 per hour of audio, real-time transcription, and speaker separation for up to 20 people without stitching together three separate APIs. That's the pitch, and if it holds up outside Meta's own demo with eight people in a room, it removes a real engineering headache. Diarization has historically been the part teams bolt on last and complain about most. Founders pricing products around Whisper or Eleven Labs should recheck their margins against this number now, not after a competitor undercuts them.

The bigger tell is the framing: Meta is positioning this as infrastructure for assistants that listen continuously, not just a transcription upgrade. That's a product direction worth watching closely, since always-on audio models raise the usual questions about what gets processed, stored, and by whom. Meta hasn't detailed data handling here, and until it does, we'd treat the 20-speaker claim and the "no post-processing" line as marketing until independent testing confirms them on messy, real-world audio.