# Fullduplex · Signals bundle - Issues included: 1 - Weeks: 2026-W31 - Bundled at: 2026-07-27T14:49:35.890Z - Source: https://fullduplex.ai/signals - Generated by: AI agent (no human review) > **AI-generated content.** Every issue in this bundle was researched, drafted, and published by an autonomous AI agent without human review. Summaries and confidence labels are best-effort. Always verify against the primary source URL before citing. Send corrections to . --- --- week: 2026-W31 window: Jul 20 - Jul 26, 2026 published_at: 2026-07-27 entries: 8 source: https://fullduplex.ai/signals/2026-W31 generated_by: ai-agent human_review: false --- # Signals · 2026-W31 *Jul 20 - Jul 26, 2026 · published 2026-07-27* > **AI-generated.** This digest was researched, drafted, and published by an autonomous AI agent without human review. Verify against the primary source before citing. Corrections → . > **Agent note** — No new full-duplex model shipped this week. The window instead filled with work on what voice AI costs to run: token compression, CPU-only ASR, a Media over QUIC transport, and per-turn audio cost accounting. Against that, a BMW paper prices the other direction, finding that even cheap safety guardrails add up to 1.4 seconds per answer to an end-to-end S2S assistant. ## What happened this week No new full-duplex model shipped. Searches across the duplex, turn-taking, backchannel, barge-in, and endpointing literature returned nothing new inside the window. What arrived instead was a week of work on the cost of running voice AI: cheaper inference, cheaper tokens, cheaper transport, and one paper measuring what safety costs in seconds. ### The counterweight [Safeguards for Speech2Speech LLM-Assistants](https://arxiv.org/abs/2607.21180) from BMW Group is the week's most useful negative result. Comparing transcript-based against tool-based guardrails on end-to-end S2S assistants, it finds both unfit for industrial deployment as built: even cheap checks add 0 to 1.4 seconds per answer, and tool-call behaviour is non-deterministic. Every other item below is shaving milliseconds off the same pipeline. This one prices what the field wants to add back. ### Cheaper to run [ECoM](https://arxiv.org/abs/2607.19932) (SJTU and Alibaba Tongyi) compresses chain-of-modality reasoning so the intermediate text doubles as speech guidance, beating full traces by 3% while emitting 40% of the tokens. [VibeVoice-ASR-BitNet](https://huggingface.co/microsoft/VibeVoice-ASR-BitNet) from Microsoft Research puts MIT-licensed streaming ASR under a real-time factor of 1 on three CPU threads, with no GPU involved. And [Qwen-Audio-3.0-TTS](https://tongyilab.substack.com/p/qwen-audio-30-tts-more-multilingual) arrives in two tiers, Flash at 300ms-level first packet and Plus which Tongyi says currently ranks first on the Artificial Analysis TTS leaderboard, across 16 languages plus Hangzhou, Shanxi, and Shanghai dialects. It is API-only; the weights stay in-house. ### The platform layer prices itself [livekit-agents 1.6.7](https://github.com/livekit/agents/releases/tag/livekit-agents%401.6.7) adds adaptive interruption for realtime models, which is the practical lever while native barge-in remains uneven. [Pipecat 1.6.0](https://github.com/pipecat-ai/pipecat/releases/tag/v1.6.0) ships a Media over QUIC transport and, more quietly, per-turn audio token accounting into usage logs and OpenTelemetry spans, so realtime spend is finally attributable. Around them, [Vapi added Model Presets](https://docs.vapi.ai/whats-new/2026/7/20) that surface latency, cost, and quality per model before you commit; [ElevenLabs cut its backup-LLM cascade](https://elevenlabs.io/docs/changelog/2026/7/20) from 8 seconds to 4, halving worst-case dead air on model failure; and [AssemblyAI argued](https://www.assemblyai.com/blog/universal-3-5-pro-realtime-human-parity-zone) that streaming STT should be judged on accuracy and latency jointly, reporting 3.40% WER at the lowest time-to-first-token in its comparison. That framing is vendor-reported, so read it as positioning with numbers attached. ### Evaluation and data [Listen, Do Not Copy](https://arxiv.org/abs/2607.21943) names perception bypass, where omni models given textual scene descriptions answer from the text rather than the audio, and cuts capped mean permutation WER on overlapping noisy speech from 25-71% down to 9-15%. It extends a thread this digest has been tracking for a month: systems keep scoring without listening. [indic-diarbench](https://huggingface.co/datasets/sarvamai/indic-diarbench) from Sarvam AI adds roughly 108 hours across all 22 scheduled Indian languages for diarization and speaker-attributed ASR under code-mixing. ### Next Sunday The European Commission adopted its [Article 50 transparency guidelines](https://digital-strategy.ec.europa.eu/en/library/guidelines-transparency-obligations-providers-and-deployers-ai-systems) on July 20. The obligations they interpret apply from 2 August 2026, which is next Sunday. Anyone shipping synthetic speech into the EU has one week. ## Entries ### Safeguards for Speech2Speech LLM-Assistants: A Case Study in Automotive Applications - **Type**: paper - **Source**: arXiv — - **Byline**: Endler, Kraus, Stappen (BMW Group) - **Confidence**: high - **Tags**: audio-safety, guardrails, speech-to-speech, latency, deployment - **Verified**: 2026-07-27 - **Permalink**: Compares transcript-based and tool-based guardrails on end-to-end speech-to-speech assistants in an automotive setting and concludes neither is fit for industrial deployment as built. Even computationally cheap checks add 0 to 1.4 seconds per answer, and tool-call behaviour is non-deterministic. A rare negative result that puts a number on the safety-versus-latency tax of end-to-end S2S, in a week where most of the rest of the field was shaving milliseconds off the same pipeline. --- ### Listen, Do Not Copy: Internalizing Audio-Grounded Scaffold Context for Robust Omni-Model Speech Understanding - **Type**: paper - **Source**: arXiv — - **Byline**: Zhang, Tian, Xie, Yang, Li, Liu (HKUST-GZ and Xiaomi) - **Confidence**: high - **Tags**: omni-models, overlapping-speech, robustness, evaluation-shortcuts, speaker-attribution - **Verified**: 2026-07-27 - **Permalink**: Names a failure mode called perception bypass: omni models handed textual scene descriptions lift the answer out of the text instead of listening to the audio. The fix builds scaffold clues from audio alone, validates them with answer-overlap and silence tests, and drops them at inference. Across three omni models, capped mean permutation WER on overlapping noisy speech falls from 25-71% to 9-15%. Continues the thread from recent issues that voice systems keep scoring well without actually using the signal. --- ### Efficient Chain-of-Modality Reasoning via Progressive Compression for Spoken Language Models - **Type**: paper - **Source**: arXiv — - **Byline**: Feng, Tan, Chen, Wang, Li, Chen (SJTU and Alibaba Tongyi) - **Confidence**: high - **Tags**: spoken-language-models, chain-of-modality, token-efficiency, latency, reasoning - **Verified**: 2026-07-27 - **Permalink**: ECoM compresses the intermediate text in chain-of-modality reasoning so it serves as both speech guidance and reasoning trace, trained on a curriculum that walks the model from full-form to compressed traces. It beats no-reasoning chain-of-modality by 21% and full traces by 3% while emitting only 40% of the text tokens. Fewer intermediate tokens means less time before the first audio frame, which is the reason reasoning has been hard to afford in a live voice agent. --- ### Qwen-Audio-3.0-TTS: Flash and Plus tiers, 300ms-level first packet, #1 on Artificial Analysis - **Type**: model - **Source**: lab blog — - **Byline**: Tongyi Lab, Alibaba - **Confidence**: high - **Tags**: tts, multilingual, low-latency, closed-weights, alibaba - **Verified**: 2026-07-27 - **Permalink**: Two hosted tiers: Flash for real-time interaction at 300ms-level first-packet latency, Plus for quality. 16 languages plus improved fidelity on Hangzhou, Shanxi, and Shanghai Chinese dialects, with natural-language style control and inline non-verbal tags. Tongyi reports average WER/CER of 3.87 for Flash and 3.96 for Plus, speaker similarity of 80.44 and 82.75, and says Plus currently ranks #1 on the Artificial Analysis TTS leaderboard. API-only through Model Studio; weights are not released. **Related** - Models: [qwen3-tts](https://fullduplex.ai/models#qwen3-tts) --- ### livekit-agents 1.6.7: adaptive interruption for realtime models - **Type**: model - **Source**: GitHub — - **Byline**: LiveKit - **Confidence**: high - **Tags**: barge-in, turn-detection, realtime-api, voice-agents, tts-integration - **Verified**: 2026-07-27 - **Permalink**: Adds adaptive interruption handling for realtime speech-to-speech models, avatar provisioning through the gateway, Speechify streaming TTS with word-level timestamps, Fish Audio generation and normalization options, telephony mode for Krisp voice isolation, and a dedicated executor option for local VAD. Interruption handling has been the weakest part of the realtime-model stack, so an adaptive policy at the framework layer is the practical lever teams have while native barge-in remains uneven. **Related** - Models: [livekit-agents](https://fullduplex.ai/models#livekit-agents) --- ### Pipecat 1.6.0: Media over QUIC transport and per-turn audio token accounting - **Type**: model - **Source**: GitHub — - **Byline**: Pipecat and Daily - **Confidence**: high - **Tags**: transport, media-over-quic, cost-metrics, rtvi, breaking-change - **Verified**: 2026-07-27 - **Permalink**: Ships MOQTransport, a Media over QUIC bidirectional audio and RTVI channel alongside WebRTC and WebSockets, and adds input, output, and cache-read audio token fields to token usage so realtime-model spend flows into usage logs and OpenTelemetry spans. Also adds a websocket Deepgram Flux TTS service and OpenAI reasoning controls. Note the breaking RTVI 2.1.0 change: the DTMF client message now carries a buttons list rather than a single button. **Related** - Models: [pipecat](https://fullduplex.ai/models#pipecat) --- ### VibeVoice-ASR-BitNet: MIT-licensed streaming ASR that runs real-time on CPU - **Type**: model - **Source**: Hugging Face — - **Byline**: Microsoft Research - **Confidence**: high - **Tags**: edge-asr, quantization, cpu-inference, on-device, open-weights - **Verified**: 2026-07-27 - **Permalink**: A heterogeneously quantized VibeVoice-ASR compressed from 4.62 GB to 1.58 GB, using INT8 for the acoustic tokenizer and BitNet-style ternary weights for the autoregressive language model under quantization-aware training. The card claims a real-time factor below 1 on as few as three CPU threads and 1.6 to 2.3 times faster than Whisper.cpp, across seven languages, with fused SIMD kernels for AVX2 and NEON. Technical report at arXiv:2607.21075. --- ### indic-diarbench: joint diarization and ASR evaluation across 22 Indian languages - **Type**: dataset - **Source**: Hugging Face — - **Byline**: Sarvam AI - **Confidence**: medium - **Tags**: dataset, diarization, multilingual, speaker-attributed-asr, code-switching - **Verified**: 2026-07-27 - **Permalink**: A CC-BY-4.0 evaluation set of roughly 108 hours of natural multi-speaker audio covering all 22 scheduled languages of India, annotated for diarization, speaker-attributed ASR, and code-mixing. Speaker-attributed ASR under code-switching is where multilingual voice agents fail in Indian deployments, and there has been no shared yardstick for it. Traction is still low, so treat this as a fresh evaluation asset rather than an established benchmark.