Wire flash
Unnamed startup beats OpenAI, ElevenLabs, Deepgram, AssemblyAI in AI speaker identification benchmark
Editorial responsibility
- No named human review is recorded for this page.
- Source reporting is collected, normalized, translated or condensed automatically when needed.
- Automatically published source-backed update
An unnamed startup has outperformed four major AI speech companies—OpenAI, ElevenLabs, Deepgram, and AssemblyAI—in a benchmark for speaker identification, a critical capability for AI notetakers and ambient assistants. The startup achieved lower error rates than every commercial system tested across office meetings, noisy meetings, and real dinner parties, finishing 7.6 points ahead of the next-best system overall. The approach involves identifying who is speaking when, recovering overlapping voices when confidently assignable, transcribing each person separately, and recognizing people using clean voice samples. The team also reports achieving this performance at a tenth of the cost of competing systems.
Source report
OpenAI. ElevenLabs. Deepgram. AssemblyAI.
A startup has beaten all four in a critical task: consistently identifying speakers across conversations — a key capability for AI notetakers and ambient assistants.
In their benchmark, the startup achieved lower error rates than every commercial system tested, including in office meetings, noisy meetings, and even real dinner parties.
Overall, the startup finished 7.6 points ahead of the next-best system.
The Approach
The method that caught attention:
- Identify who is speaking when
- Recover overlapping voices when they can be confidently assigned
- Transcribe each person separately
- Recognize people using clean voice samples
The team also reports achieving this at a tenth of the cost.
Source
DataChazNeutral / independent