LIVE
Mistral Makes Large 4 Available as Open Weights29/10/26 · Mistral AI|Gemini 4 Carbon tops the Artificial Analysis Intelligence Index v529/10/26 · Google|Mistral Large 4 now available on Hugging Face29/10/26 · Mistral AI|Google Introduces "Gemini 4 Carbon" as Next Step in Its Model Lineup28/10/26 · Google|DeepMind unveils Genie 4, a new generation of world model27/10/26 · Google DeepMind|Anthropic files IPO prospectus, with risk disclosures taking center stage26/10/26 · Anthropic|Robot policy trained only in simulation matches top real-world systems26/10/26|OpenAI unveils GPT-6.1 Terra, a model built for computer use26/10/26 · OpenAI|OpenAI revises its Model Spec with five guiding values for AI behavior29/10/26 · OpenAI|OpenAI outlines its monitoring system for catching misaligned model behavior29/10/26 · OpenAI|Hugging Face releases open models for speech recognition and decision-making on Apple devices29/10/26 · Hugging Face|Hugging Face launches Project ProtoStar to make publicly funded science AI-ready29/10/26 · Hugging Face|Mistral Makes Large 4 Available as Open Weights29/10/26 · Mistral AI|Gemini 4 Carbon tops the Artificial Analysis Intelligence Index v529/10/26 · Google|Mistral Large 4 now available on Hugging Face29/10/26 · Mistral AI|Google Introduces "Gemini 4 Carbon" as Next Step in Its Model Lineup28/10/26 · Google|DeepMind unveils Genie 4, a new generation of world model27/10/26 · Google DeepMind|Anthropic files IPO prospectus, with risk disclosures taking center stage26/10/26 · Anthropic|Robot policy trained only in simulation matches top real-world systems26/10/26|OpenAI unveils GPT-6.1 Terra, a model built for computer use26/10/26 · OpenAI|OpenAI revises its Model Spec with five guiding values for AI behavior29/10/26 · OpenAI|OpenAI outlines its monitoring system for catching misaligned model behavior29/10/26 · OpenAI|Hugging Face releases open models for speech recognition and decision-making on Apple devices29/10/26 · Hugging Face|Hugging Face launches Project ProtoStar to make publicly funded science AI-ready29/10/26 · Hugging Face|
+ nAIvigateA SHEET FOR PEOPLE WITHOUT 2 HOURS
+ nAIvigate · fiches / agents-vocaux-relais-ou-speech-to-speech
▣ fiche-011.md ×
□ index.md
☰ ContentsVerdictRelay or speech-to-speech: six criteriaCase study: Dessa, by BlockWhat the rules already requireWarning signs

Voice agents: the relay that transcribes, the model that listens.

Voice agents are changing architecture. Instead of chaining three models, new models listen and speak at the same time and keep the emotion. What gains in naturalness loses in control: there is no longer any text to filter before the agent speaks.

Verdict

the essentials in 10 seconds3 rows
NATURALModels that listen and speak at once make the voice almost human. Laughter, hesitation and irony finally come through.SHIFT
CONTROLWith no intermediate text, nothing is left to filter or review: the relay's guardrails no longer apply.DECISIVE
RULESSay the caller is talking to an AI, do not analyse employees' emotions, treat the voice as sensitive data. The rules already exist.AI ACT

The relay that transcribes, the model that listens

what is lost, what is gained2 architectures

Most voice agents chain three models: transcription, LLM, text to speech. New audio-to-audio models replace the chain with a single model that keeps the tone, but leaves no text to check.

THE RELAY: THREE MODELS IN A ROW Transcriptionspeech to text LLMwrites the reply Voice synthesistext to speech lost: tone, laughter, sighs text you can filter and log tone rebuilt, flatter + delay at every step · waits for you to finish · blocks replaceable one by one SPEECH-TO-SPEECH: ONE MODEL, FULL DUPLEX Audio-to-audio modellistens and speaks at the same time you speak it interrupts, laughs you keeps emotion and rhythm no text to filter or review → The more natural the voice, the more the guardrails must work on the audio itself

Relay or speech-to-speech: six criteria

to choose by use case, not by demo6 rows
naturalnessSpeech-to-speech keeps emotion and interrupts like a human. The relay stays more mechanical.S2S
latencyA single model avoids waiting on three steps. The relay stacks delays and waits for the end of the sentence.S2S
controlThe relay produces text that can be filtered before it is spoken. Speech-to-speech speaks directly.RELAY
logThe relay leaves a transcript the security and support teams can read. Speech-to-speech requires a separate transcription.RELAY
blocks ↗The relay lets you swap the voice, the transcription or the LLM separately, including for a European model.RELAY
maturityThe relay is proven in production. The best speech-to-speech models are recent, and some are not yet available.RELAY

Case study: Dessa, by Block

announced in October 20265 rows
the claimAn audio-to-audio model trained only on human conversations, with no text and no pretrained language model.
what it doesFull duplex: it listens while it speaks, recognises irony, frustration or laughter, and expresses them in turn.
localJack Dorsey says it runs on a recent MacBook Pro, offline. A key point for call privacy.WATCH
availabilityNo API, no pricing, no published weights: a waitlist and a small group of safety researchers.NOT YET
to verifyParameters and benchmarks unpublished: all performance claims come from Block, no independent evaluation so far.CAUTION

What the rules already require

before deploying a voice agent5 rows
transparency ↗Article 50 of the AI Act requires telling people they are interacting with an AI. All the more when the voice sounds human.ART. 50
emotionsEmotion recognition is banned in the workplace and in education by Article 5 of the AI Act, except for medical or safety reasons.BANNED
voiceWhen used to identify someone, the voice becomes biometric data under the GDPR, with the stronger safeguards that come with it.RGPD
cloningCloning a voice requires the person's written consent, and use limited to what they agreed to.
fraudSynthetic voices make CEO fraud more convincing. Set up a code word or a call-back to a known number for any transfer requested by phone.PRIORITY

Warning signs

one is enough to dig deeper6 rows
disclosureThe agent never says it is an AI, or only when asked.BLOCKER
emotionsThe vendor sells an "emotion score" for employees or candidates.BANNED
recordingCalls are kept and reused for training with no clear duration or agreement.
transcriptNo transcript kept: no way to know what the agent promised a customer.
actionsThe voice agent can change an account or approve a payment on the strength of the voice alone.BLOCKER
tools ↗Compare the voice, transcription and cloning tools rated on nAIvigate.DIRECTORY
A voice agent in your customer service? We help you choose the architecture, place the guardrails and follow the rules, from transparency to voice data.
TALK TO nAIvigate STUDIO →
⎇ mainfiche #011 · securitesources: Block, annonce de Dessa, octobre 2026 (performances autodéclarées) · Règlement (UE) 2024/1689 (AI Act), articles 5 et 50 · RGPD, article 9 (données biométriques) · Synthèse nAIvigate, octobre 2026
Follow nAIvigate | naivigate.com/en/fiches