What happened this week
Four preprints worth forwarding, all evaluation-leaning. The Hugging Face / GitHub / lab-blog buckets did not surface a primary-sourced release in scope for this window, so the issue is paper-only.
The headline — full-duplex evaluation
HumDial-FDBench is the comprehensive write-up of the ICASSP 2026 HumDial Challenge full-duplex track. The headline contribution is a dual-channel dataset of real human-recorded conversations — capturing interruptions, overlap, and feedback mechanisms — and a public leaderboard that compares open-source and proprietary systems on interruption handling and conversational flow. This is the most concrete shared eval artifact full-duplex has had since FD-Bench v3, and it slots in as a peer to the existing HumDial track Fullduplex already tracks under /benchmarks.
Paralinguistic and timing control
Two papers push the controllability frontier rather than capability:
- SpeechParaling-Bench expands paralinguistic feature coverage from fewer than 50 to over 100 fine-grained features, with 1,000+ English-Chinese parallel queries across three tasks (fine-grained control, intra-utterance variation, context-aware adaptation). The pairwise-comparison evaluation pipeline is the methodologically interesting bit. The headline empirical finding is that paralinguistic misinterpretation accounts for 43.3 percent of errors in situational dialogue even on leading proprietary models.
- MAGIC-TTS is presented as the first TTS system with explicit token-level local timing control over both content duration and pause. The training mechanisms — high-confidence duration supervision plus zero-value bias correction — are the parts that read as transferable. The scenario-based editing benchmark covers navigation guidance, guided reading, and accessibility-oriented code reading.
Fairness and robustness in ASR
Do LLM Decoders Listen Fairly? ships a 216-run stress test of nine ASR models across three architectural generations (CTC-only, encoder-decoder, and explicit-LLM decoder) on Common Voice 24 plus Meta's Fair-Speech. The two findings worth reading are that LLM decoders do not amplify racial bias — Granite-8B has the best ethnicity fairness in the sweep — and that audio compression, not LLM scale, is the dominant predictor of accent fairness. Whisper enters catastrophic repetition loops under chunk masking, while explicit-LLM decoders produce ~38x fewer insertions.
What is not here
No Hugging Face / GitHub / lab-blog signal landed inside the window with a primary source we can cite. The Mistral, Hume, and Microsoft speech releases that sometimes get cited under "this week" are all earlier in March or early April. If something open-weights ships before next Sunday, it will move into 2026-W19.
Corrections to hi@fullduplex.ai.