What happened this week
Two new benchmarks measure how often voice agents finish real work, and both numbers are low. Meanwhile, the vendors that sell the outer parts of a cascaded agent, the speech recognizer (automatic speech recognition, ASR) and the voice (text-to-speech, TTS), shipped faster and more accurate versions. Those releases improve how well an agent hears a caller and how quickly it answers. Neither benchmark finds its main failures there.
Voice agents and real work
APEX-Voice, from the University of Maryland, gives voice agents 120 professional workflows across ten kinds of work, including form completion and interviewing. Each runs in a simulated workplace with typed tools and authorization rules; a run passes only if the agent reaches the right end state through a valid process, completes the actions and produces a valid final document. Across GPT-Live-1, Gemini-3.8-Live, Grok-Voice-Think-2.0, Step-Audio3 and GPT-realtime-2.1, none passes more than 25% of workflows on the first attempt (Pass@1), and the best system passes all three repeats of a workflow (Reliable@3) only 10.8% of the time. Keeping state straight across steps is the most common failure, and success falls further when the caller corrects themselves mid-sentence.
VAmoS Part Deux, from Veris AI, runs 100 utility-billing calls in which each caller makes two to four requests. The agent works through sixteen tools against a simulated billing system, and account access stays locked until the caller is verified. Across fourteen voice stacks, completion ranges from 17.3% to 44.7%. Grok Voice leads, and Gemini 3.8 Live and GPT-Live follow at about the same cost per call. A television in the background cuts pooled completion from 38.7% to 8.6%. The simulated caller often accepts a wrong result because it hears what the agent says but cannot inspect what it changed, so a score built from the conversation alone overstates an agent.
DSB-DG, from Adobe Research and the University of Maryland, finds that open-weight speech-to-speech models collapse abruptly as a supplied document grows, and fail more often by making answers up than by declining. VoxParity keeps the transcript fixed and changes only the audio, such as a medical monitor beeping or a frightened whisper, so the correct tool call changes with it. Only 11 of 23 testable systems act on what they hear more than a transcript-only pipeline does.
More than two speakers, and when to speak
MultiTalk, from CUHK MMLab and accepted at NeurIPS 2026, extends the Moshi design to long conversations among several people in English and Chinese. It releases 57.6k hours of synthetic training data and MultiTalkBench, built from real recordings that average 32.6 minutes. The authors report that their model substantially outperforms Moshi, MiniCPM-o-4.5 and Qwen3-Omni-30B-A3B-Instruct on that benchmark.
DuplexAct-Bench scores 12 full-duplex systems on 1,290 English and Chinese trials that cover six behaviours, among them interrupting, yielding, deliberate silence and backchanneling (the short "mm-hm" a listener gives without taking the turn). A good reply often arrives at the wrong moment. Fullduplex is published by oto, whose otoSpeech-full-duplex-280h dataset supplies the natural conversations in one of the benchmark's backchannel subsets.
Faster parts for cascaded agents
A cascaded agent chains three models: ASR turns the caller's audio into text, a language model writes the reply, and TTS speaks it. Three vendors upgraded the first and last links this week. All figures below are vendor-reported.
AssemblyAI Universal-3.6 Pro Realtime costs $0.45 an hour and covers 32 languages. On AssemblyAI's own benchmark of 12,460 scripted calls, its word error rate (WER) is 5.19%, against 7.78% for ElevenLabs Scribe v2 and 8.64% for Deepgram Nova-3. Two results bear on turn-taking. Short replies such as "yes" are wrong 1.45% of the time, down from 2.65%. On 2,950 clips of hold music, line noise and silence, it returns text 0.27% of the time, against 38.75% for Deepgram Flux. Phantom text on hold music is what makes an agent start talking to nobody.
Microsoft AI released MAI-Transcribe-2-Streaming, a 60-language streaming recognizer whose first partial transcript arrives "in just over 100ms", at an introductory $0.54 an hour. It came with two voices: MAI-Voice-2.1 at $22 per million characters and MAI-Voice-2.1-Flash at $15 with 150 ms end-to-end latency. ElevenLabs released Eleven v4 and Eleven v4 Turbo, reporting about 100 ms median inference and about 150 ms median time to first speech for Turbo. Turbo runs over the Text to Dialogue WebSocket, so existing integrations need a new endpoint.
Full-duplex video
Tavus announced Griffin-Lite, a research preview that runs a video call as one full-duplex system. It takes in the caller's audio and video, decides at sub-second intervals whether to speak, react or wait, and generates 720p video of a face. NVIDIA's VideoFDB leaderboard, which uses a language-model judge, scores it 3.83 out of 5 for generation against a human reference of 3.92, and 3.73 for perception against 4.20. Its median response there is 1,892 ms and 2,232 ms, against 900 ms and 1,400 ms for the humans. In Tavus's own one-minute call study, 26 of 54 participants (48%) took it for a real person. Only selected testers have access.
Open releases
Bilibili's IndexTeam released Index-Echo-S2ST in 9B and 2B sizes under Apache-2.0. It dubs Chinese or English speech into English, Spanish, Japanese or Chinese in the source speaker's voice. Against a matched two-step pipeline, the 9B end-to-end path has lower content error in four of six directions; Chinese to English favours the pipeline.
Hume closes EVI
Hume will shut down its TTS and Empathic Voice Interface (EVI) APIs on November 13, 2026, and will delete account data afterwards. EVI was one of the first commercial speech-to-speech APIs built around emotional expression. Hume gave no reason.
TurnBench
Since last week's issue, an external entry, FD-VAD, joined TurnBench in ninth place: 0.826 recall, 0.049 false-positive rate, 1,025 ms. It matches ESPnet Turntaking on recall with fewer false triggers, but is 163 ms slower. We found no write-up, which submissions must now include. Fullduplex is published by oto, which supplied TurnBench's training split and has co-authors on the paper.