Voice AI Model Comparison
A side-by-side of the full-duplex / speech-to-speech voice models security and governance teams are now being asked about — Block Dessa, OpenAI Realtime (Advanced Voice), Kyutai Moshi, Sesame CSM, Hume EVI, ElevenLabs Conversational and Google Gemini Live. Compare them on architecture (native speech-to-speech vs STT→LLM→TTS cascade), full-duplex support, hosting (local vs cloud), open weights and availability — with a security-relevant note on each, because where audio runs decides where your data goes. Filter to what you need; every fact is vendor-stated as of October 2026 and should be verified.
Facts as of October 2026, vendor-stated and necessarily hedged. This is a fast-moving category — open weights, APIs, benchmarks and availability change constantly. Verify against each vendor before relying on any cell. No benchmarks or performance claims are invented here; where a vendor has published nothing, the cell says so.
| Model | Architecture | Full-duplex | Hosting | Open weights | Availability |
|---|---|---|---|---|---|
| DessaBlock | Speech-to-speech (direct audio in/out) | FullFull-duplex (interrupts, handles pauses) | Local (runs on an M4 Max MacBook, per Block) | No public weights | ResearchResearch preview (AI-safety researchers) |
| Realtime / Advanced VoiceOpenAI | Speech-to-speech (Realtime API) | FullFull-duplex streaming | Cloud API | No (proprietary) | GAGenerally available (API) |
| MoshiKyutai | Speech-to-speech (full-duplex) | FullFull-duplex (research design goal) | Self-host or cloud (open release) | Yes (open release) | GAOpen research release |
| CSMSesame | Conversational speech model | PartialConversational; full-duplex extent varies | Cloud product; components opened | Partial (some components open) | PreviewPreview / limited |
| EVIHume | Empathic voice interface (cascade + prosody) | PartialLow-latency turn-taking | Cloud API | No (proprietary) | GAGenerally available (API) |
| Conversational AIElevenLabs | Orchestrated STT → LLM → TTS cascade | PartialManaged turn-taking / interruption | Cloud platform | No (proprietary) | GAGenerally available |
| Gemini LiveGoogle | Multimodal live (native audio) | FullFull-duplex live streaming | Cloud API (Google Cloud / AI Studio) | No (proprietary) | PreviewPreview / rolling GA by surface |
Security-relevant notes
Vendor-stated local inference means audio can stay on-device; no public weights, API, paper or benchmarks yet, so every claim is Block’s own and unverified. Announced 2026-10-23; trained on speech alone (no text / pretrained models).
Cloud-hosted: audio and transcripts leave your environment to a third party, so data-residency, retention and recording-consent terms govern. Check the enterprise data-use and retention commitments before sending PII-bearing calls.
Open weights allow fully self-hosted, on-prem deployment — audio need never leave your network. The trade-off is you own the hardening, patching and abuse controls; nothing is managed for you.
Mixed posture: some components have been opened while the product is cloud-delivered. Confirm exactly what you can self-host vs what routes to the vendor before assuming audio stays local.
Cloud API that also infers emotional/prosodic signals from the voice — that derived affect data is itself sensitive. Treat emotion inference under the same consent and retention rules as the raw audio.
Cascade routes audio through separate recognition, reasoning and synthesis stages — more components, more places data transits and is logged. The platform also offers voice cloning, so verify identity and consent controls for any synthesised voice.
Cloud-hosted on Google’s platform: governed by your Google Cloud data-processing, region and retention settings. Availability and exact capabilities differ by surface (API vs app), so pin to the specific product terms you deploy under.
How to read this security-first
- Speech-to-speech vs cascade. A native speech-to-speech model processes audio directly; a cascade runs speech‑to‑text → an LLM → text‑to‑speech. The cascade has more components and therefore more places audio is transcribed, logged and can leak — but its text middle is also easier to filter and audit.
- Local vs cloud is the data-residency lever. Local / self-hosted inference means audio can stay on the device or in your network; a cloud API means audio and any transcript leave your organisation to a third party you must have terms with.
- Open weights cut both ways. They enable fully on-prem deployment and independent review, but shift all the hardening, patching and abuse-control burden onto you — a managed cloud service does that part for you.
Educational — not affiliated with Block, OpenAI, Kyutai, Sesame, Hume, ElevenLabs, Google or any vendor listed. All capability, hosting and availability facts are vendor-stated as of October 2026 and change frequently; verify against each vendor’s current documentation. Runs entirely in your browser: nothing is uploaded and no audio is captured.
What this comparison covers
- A filterable, security-first comparison of full-duplex / speech-to-speech voice AI models — Block Dessa, OpenAI Realtime (Advanced Voice), Kyutai Moshi, Sesame CSM, Hume EVI, ElevenLabs Conversational and Google Gemini Live.
- Columns cover the choices that carry security weight: architecture (native speech-to-speech vs STT→LLM→TTS cascade), full-duplex support, hosting (local / cloud / both), open weights, availability (GA / preview / research-only) and a security-relevant note per model.
- Block announced Dessa on 2026-10-23 as a speech-to-speech, full-duplex research preview that runs locally on an M4 Max MacBook and was trained on speech alone — with no public weights, API, paper or benchmarks yet, so its claims are Block’s own and unverified.
- Filter by whether a model can run locally, has open weights, or is generally available now. Facts are vendor-stated as of October 2026 and change constantly — verify before relying.
- Runs entirely in your browser. No account, no audio captured, nothing uploaded.
Why this matters
Full-duplex speech-to-speech voice models change the risk picture in a way the older text-chatbot era did not. The audio itself is the input and the output, so there may be no text transcript to filter, log or audit in the middle; the conversation can be interrupted and steered in real time; and the person on the other end may not realise they are talking to a machine. Block’s Dessa announcement on 2026-10-23 — a research preview that runs locally and was trained on speech alone — is one marker of how fast the category is moving, but it sits alongside cloud offerings from OpenAI, Google, Hume and ElevenLabs and the open Moshi release from Kyutai.
For a security leader the first questions are not about voice quality: they are where the audio goes (local vs cloud), whether you can audit what the model does, how you verify the caller, and what happens to recordings. This comparison keeps those trade-offs in front of you. Every claim here is the vendor’s own as of October 2026 — and for Dessa in particular there are no public weights, API, paper or benchmarks yet, so treat its stated capabilities as unverified until independent evidence exists.
Frequently asked questions
What is a speech-to-speech (full-duplex) voice AI model?
A speech-to-speech model processes and generates audio directly, without a separate speech-to-text step in the middle. "Full-duplex" means it can listen and speak at the same time — it can be interrupted mid-sentence and handle natural pauses and overlaps, rather than taking strict turns. This contrasts with the older cascade or pipeline approach, which chains speech-to-text, then a text LLM, then text-to-speech. Full-duplex speech-to-speech feels more like a natural conversation; the cascade is easier to filter and audit because it exposes a text transcript in the middle.
What is Block Dessa and what is actually known about it?
Dessa is a voice AI model announced by Block (Jack Dorsey’s fintech company) on 2026-10-23. Block describes it as a speech-to-speech, full-duplex model that processes and generates audio directly with no speech-to-text step, can interrupt and handle pauses, runs locally on an M4 Max MacBook, and was trained on speech alone — no text and no pretrained models. It is a research preview with early access limited to AI-safety researchers. As of October 2026 there are no public weights, no API, no paper and no published benchmarks, so every performance characteristic is Block’s own unverified claim. Treat it accordingly until independent evidence appears.
Which voice AI models can run locally versus in the cloud?
Local or self-hostable options let audio stay on the device or inside your network. Block states Dessa runs locally on an M4 Max MacBook (research preview). Kyutai Moshi is released openly and can be self-hosted. Sesame has opened some components. OpenAI Realtime / Advanced Voice, Hume EVI, ElevenLabs Conversational and Google Gemini Live are cloud APIs, where audio and any transcript leave your organisation to a third party you must have data terms with. Use the "can run locally" filter in the tool to narrow to on-device / self-hosted candidates.
Why does the architecture matter for security?
Where the model runs and how it is built decide where your audio goes and what can be inspected. A cloud API means audio and transcripts leave your environment, so data residency, retention and recording-consent terms govern. A cascade (STT→LLM→TTS) has more components — more places audio is transcribed and logged — but its text middle is easier to content-filter and audit. A native speech-to-speech model has no text checkpoint, so prompt-injection and jailbreak defenses must work on audio directly. None of these is "most secure"; the right choice depends on your data-residency needs, your ability to operate the controls, and your threat model.
Does this comparison connect to any service or capture audio?
No. It is a static, in-browser comparison table you can filter. It makes no network calls, captures no audio, and takes no account or input. All facts shown are vendor-stated as of October 2026 and may be out of date — verify against each vendor’s current documentation before relying on them.
This tool is free and always will be. Would a Pro workspace — saved designs, team sharing and exports — be worth paying for?
Show me Pro →No account needed to use this tool — this just tells us whether Pro is worth building out.