In plain words: A new score checks whether the same person keeps one speaker identity across meetings, instead of renumbering speakers separately in each recording. Under this stricter test, their end-to-end system scored lowest at 47.13 versus 54.75 for the next system, and the commercial ranking changed.
Abstract · Who Said What, and Will It Be Remembered? Evaluating Persistent Speaker Attribution Across Meetings
Speech transcripts used as long-term memory must preserve both words and stable speaker identities. Existing meeting-transcription metrics either ignore speakers or remap anonymous speakers independently in each recording, so they cannot measure whether the same person retains one identity across meetings. We evaluate persistent speaker attribution with Speaker Identified cpWER (SI-cpWER), which scores a corpus under one global speaker-ID assignment. The benchmark covers five commercial diarize-then-identify cascades, two open academic baselines, and ThyVoice on the full 129-meeting CHiME-8 NOTSOFAR evaluation set in clean and noiseaugmented form, plus CHiME-6. ThyVoice is our end-to-end reference system; it repairs overlap and gates the evidence used to create and update voiceprints. Requiring persistent identity changes the commercial ranking: ThyVoice records lower SI-cpWER than every evaluated commercial cascade in all three conditions and the lowest mean in the full panel, 47.13 versus 54.75 for the next system. Complementary lexical, diarization, per-recording attribution, and speaker-clustering diagnostics characterize upstream error surfaces in the final attributed record. These results show why persistent attribution must be evaluated directly in systems that reuse conversations across time.
Shantanu Vispute, Aditya Mishra, Siddhartha Saxena
arXiv:2609.39344 · cs.SD, cs.AI · submitted Sep 30, 2026
abstract · pdf · html · A short version is accepted at IEEE SLT 2026, Demo Track
ThyVoice outperforms the top 8 speech systems, including those built with OpenAI, ElevenLabs, Deepgram, AssemblyAI, and pyannoteAI, as well as academic baselines.
"Who said what" sounds simple, but it's actually two hard problems stacked together:
- Diarization: splitting a recording by speaker - Speaker ID: recognizing the same person across conversations
If either one fails, the result is incorrect attribution and polluted data.
Our contribution is twofold. First, we introduce a new metric, Speaker-Identified cpWER (SI-cpWER). Second, we present ThyVoice, which achieves the best SI-cpWER on CHiME-8 NOTSOFAR and CHiME-6, benchmarks that simulate real-world meetings and noisy environments.
The key insight of our work is that pooling all samples of the same speaker across a recording performs better than using embeddings from the largest single sub-sample. We also present a larger system that orchestrates this as a multi-step process: removing overlapping speech, stitching together each speaker's audio, and matching it against a speaker bank.
More details are in the linked arxiv pre-print.