What happened this week
Deepgram is buying Daily, the company behind the open-source Pipecat voice agent framework, and Rime launched a speech-to-speech model with a readable text layer between hearing and speaking. In research, a second paper in two weeks measures how open full-duplex models (models that listen and speak at the same time) fall apart a few minutes into a call.
Long conversations
Diagnosing and Mitigating Long-Session Degradation, from Seoul National University and the University of Seoul, runs 90 simulated 10-minute English conversations against four open models. Moshi's response timeouts jump from 12.7% at 7 to 8 minutes to 70.2% at 8 to 9 minutes, and 96.7% of its sessions have ended by minute 10. PersonaPlex and Raon-SpeechChat keep talking while the speech breaks down, reaching a final-minute word error rate (WER) of 97.9% and 98.7%.
The authors train a small LoRA adapter, extra weights added to a frozen model, at about 0.1% of each backbone. For Raon-SpeechChat, final-minute WER falls from 98.7% to 23.4% and conversation quality rises from 1.03 to 3.68 out of 5. A second change keeps recent audio in full detail and compresses older history, cutting growth of the key-value (KV) cache, the model's running memory of the conversation, by 91.7 to 95.8%. It runs slower than the base PersonaPlex at every measured point, because that model already caps its cache. Last week's FD-LHB fixes needed no retraining; these need a little. A team running one of these models should test a full 10-minute call before trusting a short demo.
Interruptions and backchannels
A full-duplex model should stop for "wait, what about the price?" and keep going through "uh-huh", which is called a backchannel. Window Likelihood, from Nanyang Technological University and Tencent, finds two failures when reinforcement learning samples a timing reward over a whole rollout, including models that talk into silence or stop answering. Its replacement computes the reward directly over the model's stop decisions. On HumDial-FDBench, which uses real recorded interruptions in two languages, F1 (a combined score for stopping and holding) rises 13 to 62 points over the base on all six open models. PersonaPlex goes from 61.7 to 77.2 against 52.4 for Kyutai's released checkpoint, which was trained on different data and a different reward. That checkpoint restarts talking into silence in 100% of tested dialogues and Window Likelihood in 0%.
Pipecat released a backchannel detector that makes the same call ahead of the transcript, at 734k parameters and about 0.3 ms of CPU per 40 ms chunk. On held-out conversations it catches 82.9% and 81.2% of interruptions within 700 ms and talks through 86.5% and 75.2% of backchannels. In practice, a bot can keep its sentence going through "right" and still stop for "no". It is English only and untested on real phone codecs or on people talking to a bot. Fullduplex is published by oto, whose otoSpeech-full-duplex-280h is in this model's training, validation and test data.
Models and platforms
Daily announced on October 29 that Deepgram is acquiring it and that Pipecat will move under an independent foundation, with a steering committee that includes Deepgram's competitors. Deepgram is adopting Pipecat as the core of its new voice AI products. The post gives no price, charter or committee names, so a team building on Pipecat with a rival vendor's speech recognition has a promise of neutrality to hold, with no governance text to read yet.
Rime launched Prism on October 28. The model writes out what it will say before speaking, so a rule can read or rewrite the text first, and Rime says it is still hearing the user while it talks. Concretely, a check could catch a wrong account number before a caller hears it. Rime ran Prism itself on the Artificial Analysis harness with no reasoning step and reports a τ-Voice score of 56.6, or 82.5 in telecom against 45.6 for GPT-Live-1, with 0.57 s to first audio. The price is $0.080 per minute.
ServiceNow published the technical report for Apriel Voice 27B Preview, MIT-licensed weights that connect a Whisper encoder to a Gemma 4 backbone. In thinking mode it scores 98.90 on Big Bench Audio and 65.51% on τ-Voice, with 1.91 s mean latency. It outputs text and tool calls, so it needs a separate synthesizer and is not a full-duplex model. Its τ-Voice figure and Prism's come from separate self-run tests, one with reasoning and one without, so they are not a direct ranking.
Liquid AI released the LFM3 family on October 27. LFM3-Omni-2B takes text, images, video and audio at 2.17B parameters and answers in text or speech, for on-device use. Its model card does not mention interruptions. Maya 2 Native, from Maya Research and new to Hugging Face, is Apache-2.0 streaming text-to-speech (TTS) for 11 Indian languages and English. Maya reports 3.49% character error rate on its main judge and second place for Hindi on VoiceArena, with overlapping rank intervals.
Other papers
Heard Enough to Act builds 786 hours of synthetic dialogue marking the earliest moment a tool call is possible. It starts correct calls an estimated 1.8 to 2.0 s early, though macro-F1 for the firing decision is only 0.48. Uniphore shows that end-of-turn detectors trained on conversation beat the same model trained on single-turn commands by up to 0.45 F1, a gap hidden when testing only on single-turn speech. A speaker impersonation attack rebuilds a voice from 3 seconds of leaked Moshi tokens at 0.78 speaker similarity, so sending tokens in place of audio does not hide who is speaking.
Also this week
Pipecat 1.15.0 lets a classifier decide when a user's turn is complete. LiveKit Agents 1.8.10 adds Reson8 and Voxtral 3, with breaking Mistral changes. The OpenAI, Gemini and xAI changelogs show no voice API change in the window.
On Hugging Face, the week was small CPU TTS (Paradee-8M v1.2, PiVo-1.4M, Leokani 7M, Delibes-8M, Silencio-2), with Indic languages and Castilian Spanish appearing. Licences leaned Apache-2.0, with CC BY-NC 4.0 on Leokani.