Block's Dessa demo, published October 23: mid-conversation, the model chimes in a "yeah" on the right beat, interrupts, and never pretends not to hear you when you cut in. Block's spec: "audio in, audio out" -- trained from scratch on human conversation recordings, no text, no pretrained backbone (official). A day later TechCrunch reached for two words: eerily human (second-hand).
It sounds like the iPhone moment of voice. But the storefront selling "full-duplex voice" these three months mixes two goods: true full-duplex speech-to-speech (S2S) listening and speaking on one stream, and text-mediated realtime TTS that obeys instructions but cannot catch your half-way interjection. Few niches widen the marketing-versus-architecture gap like this one. Six products, one table, every cell sourced.
A scope note: the October 15 cloning comparison covers turning a voice into yours; the October 21 Ming-3.0 SOP covers text-to-speech; the Grok Voice 3.0 hotspot covers single launches. This article handles full-duplex realtime conversation -- six contenders, every number sourced.
First, split the routes: realtime TTS is not full-duplex conversation
Strictly, full duplex means both directions at once: listening while generating, tolerating interruption, yielding, barge-ins. That takes understanding and generation directly on the audio stream: end-to-end S2S. A text-mediated pipeline runs recognition, response generation, and synthesis in series; at any instant only one direction is alive.
The mediated route can also be fast and interruptible, its latency numbers often better. Block's official critique targets exactly this: "turning your voice into text throws away what text cannot carry" (official) -- pauses, laughter, half-finished words die at transcription. Drama 3 and v4 Turbo belong to this route; their strengths are real, but calling them full-duplex is a sorting error.
Six on one table
| Dimension | Block Dessa | Fish Audio Drama 3 | Eleven v4 Turbo | Qwen-Audio-3.0-Realtime | NVIDIA VoiceChat 11B | Doubao Realtime Voice 3.0 |
|---|---|---|---|---|---|---|
| Duplex method | Native S2S, listens while speaking (official claim) | Not duplex, text-mediated TTS (official) | Not duplex, streaming TTS, interruption in product layer (official) | End-to-end S2S plus duplex-control submodel (official) | One streaming network, both ways (official) | Native full-duplex end-to-end (one official line, rest hearsay) |
| Text mediation | No, no-text is the pitch (official claim) | Yes (official) | Yes (official) | Undisclosed (official, vague) | Speech tokens direct (official) | No public figure |
| Latency | Faster than realtime, no number (official) | First audio 88.9ms, H200 single stream (official) | Median ~100ms, first audio ~150ms, excl. network (official) | "Milliseconds," no number (official) | Turn-taking 448ms, interrupt yield 480ms (official) | 210ms (rumor, second-hand) |
| Controllability | Emotion claimed native (official claim) | Inline director tags, per-word tone, 16 speakers (official) | Inline tags plus effects, docs admit imperfect (official) | Dynamic tone, pace, emotion, voice cloning (official) | No public figure | No public figure |
| License | No weights, no API (official) | Open weights, Community v1.0, free under $500k revenue (official) | Closed, API only | Closed, Bailian API | Open weights, research-only license (second-hand name) | "Open source" unverified (rumor, second-hand) |
| Deployment | MacBook M4 Max and up (official claim) | 48GB-class VRAM floor (second-hand inference) | None, cloud API | None, cloud API | At least 80GB VRAM (official) | No public figure |
| Price | Undisclosed | $50 per M characters, launch $10 until Nov 12 (official) | $0.04 per 1k characters (official) | Flash audio input RMB 6 per M tokens (official) | None, research (official) | Undisclosed |
| Languages | No public figure | 83 (official tags) | 91 (official page) | No public figure | No public figure | No public figure |
| Third-party validation | On no leaderboard (second-hand) | AA pronunciation 93.2% first, arena tenth (third-party) | AA preference arena leader (third-party) | AA S2S top, officially cited (official-cited) | AA open duplex top three, two boards (third-party) | None (second-hand) |
Block Dessa: best conversationalist, also unbuyable
Dessa comes from a small lab inside Block, run by alumni of Toronto's former Dessa, reporting to Jack Dorsey (official). The official list: trained on one GB200 NVL72 rack, inference faster than realtime, local on a MacBook Pro M4 Max or better, natively expressing emotion, sarcasm, laughter, sighs (official claim).
Cold water: no public weights, no API, no license; early access limited to a few AI safety researchers (official). No official benchmark; it sits on no leaderboard (second-hand). TechCrunch flagged the question: is the emotion native, or still tag-driven (second-hand)? Prefix every capability with "Block says."
Fish Audio Drama 3: open weights, not open duplex
Drama 3 launched October 13 with an official blog and arXiv report (official). The spec is serious: 35B MoE, 8 routed experts plus 1 shared, about 3B active parameters per token (official). Its real identity is "directable TTS": bracketed director tags, per-word tone edits, [sighs] tags, 16 speakers, multi-take revision; mid-sentence voice switching and single-word retakes officially confirmed (official). But all of it happens in a you-supply-text, it-returns-audio context. Full duplex it is not.
The open-source reality is "open weights, not fully open": Fish Audio Community License v1.0 -- research free, commercial free under $500k annual revenue, past that or any hosted service a commercial license, cloning real voices without consent barred (official). Latency: first audio median 88.9ms single-stream on H200 (official). Price $50 per million characters, launch $10 until Nov 12 (official); a $52M seed round closed in July (official, sources agree). User counts conflict: one third-party report says 8 million, the company's own figure tops 15 million -- not adjudicated. Third-party results split too: first on AA pronunciation robustness at 93.2%, ahead of v4's 91.7%; only tenth on the AA preference arena (third-party).
Repository: fishaudio/fish-speech, plain text github.com/fishaudio/fish-speech (GitHub API 2026-10-27: 200).
Eleven v4 Turbo: officially 100ms, still a pipe
v4 Turbo shipped September 28 alongside v4 (official). Official benchmarks: median inference around 100ms, first audio around 150ms, both excluding application and network latency, against Sonic 3.6's 262ms and GPT-4o mini TTS's 814ms (official). The models page lists 91 languages, same-voice mid-call switching (official). Official line, not a rumor -- but it excludes the two least controllable stretches. Independent tests: inference near 118ms, first audio near 165ms (second-hand).
Classification matches Drama 3: text-mediated TTS over a bidirectional WebSocket, interruption at the ElevenAgents product layer -- not full-duplex S2S (official). Closed; API $0.04 per thousand characters, subscriptions from $6 monthly (official). Coldest water: ElevenLabs' own docs admit tag adherence imperfect, and SSML unsupported (official). Block's critique of "voice tags driving emotion" went unanswered.
Qwen-Audio-3.0-Realtime: the practical route for Chinese users
Launched in July (second-hand dating); Bailian docs updated September 28 (official). It is an end-to-end realtime speech model pitching "listen-and-think" -- speech input and semantic reasoning in parallel. A built-in duplex-control submodel uses audio, semantics, and voiceprint to decide when to take the turn, resists false interruptions and locks onto the primary speaker; turn modes: server_vad, smart_turn, push-to-talk (official docs).
For Chinese readers it is the most practical of the six: closed API, but official RMB pricing -- Flash, China region: audio input 6 RMB per million tokens, text input 1.5, text output 4.5, text-plus-audio 12 (official docs) -- Chinese docs, direct domestic access. Weaknesses: latency only "millisecond-level," no number (official); no public language count; the AA S2S top rank is officially cited, not independently verified (official-cited).
NVIDIA VoiceChat 11B: the only downloadable true duplex, at 80GB
The hardest spec in the open camp: an 11B end-to-end S2S model replacing the cascade with one streaming network -- Fast Conformer encoder, Nemotron Nano v2 9B backbone blending Mamba and Transformer, TTS decoder at 22.05kHz; a side RNN-T channel transcribes the user, and a separate tool-calling channel keeps scripts out of the voice stream (official model card, relayed, two sources agree).
The duplex numbers, from Full-Duplex-Bench 1.0: turn-taking 448ms, yielding about 480ms after interruption, takeover rate 1.00, interruption quality 4.33/5 by GPT-4o (official). Big Bench Audio 38.8% is the best open full-duplex score, but the closed leading tier sits above 80 (third-party); conversational dynamics 77.8%, second among open models behind PersonaPlex at 91.0% (third-party). First open full-duplex with realtime tool calling, but the self-reported defects are many: 2-minute context cap, multi-turn degradation, runaway monologues; BFCL-v3 56.1%, tool parameter correctness only 44.2%, cap 5 tools per session (official).
The release date splits: HuggingFace weights August 3, NGC container page July 21, an August 9 account also circulates -- not adjudicated (second-hand plus official page). The license is research-only (labeled OpenMDW v1.1 by TMT Media, second-hand license name), deployment demands at least one 80GB GPU: A100, H100, RTX 6000 Pro, or B200 (official). Open, yes. Commercially usable, no. Runnable on your desk, also no.
Doubao Realtime Voice 3.0: one official line, everything else hearsay
The least transparent of the six. Reliable first-party material: one sentence in Volcano Engine's October 20 release notes -- "Added support for Doubao Realtime Voice Model 3.0 (Seeduplex), providing full-duplex end-to-end realtime voice conversation with visual understanding" (official). That confirms the full-duplex positioning, nothing else.
Every circulating number needs a rumor tag: interruption accuracy 97.3%, latency 210ms, completion 94%, from a WeChat research account (rumor, second-hand); "Seeduplex open-sourced April 2026" equally unverified, no repository or license text located (rumor, second-hand); pricing undisclosed. Until official documentation arrives, this review endorses none of it.
The open-source bind: too heavy or too old
The open landscape of 2026 fits three sentences. Dessa, the most talked-about, is fully closed. VoiceChat 11B released weights, but a research-only license plus an 80GB VRAM floor push it into datacenter territory. The only open full-duplex running on 24GB or less is still Moshi -- Kyutai's 7B from 2024, CC-BY 4.0 weights, about 200ms on an RTX 4090 (community measurements), but English and French only, a 5-minute sliding context, no tool calling (third-party wiki). Weights you cannot afford to run; runnable ones out of date.
Repository: kyutai-labs/moshi, plain text github.com/kyutai-labs/moshi (GitHub API 2026-10-27: 200).
One more cut from the boards: conversational dynamics -- PersonaPlex 91.0%, VoiceChat 77.8%, Moshi 4.3% (third-party). Freeze-Omni and FLM-Audio on the boards have unverifiable hardware requirements (second-hand); check before citing.
Do not race 88.9ms against 448ms
The latency row is easiest to misread. Drama 3's 88.9ms first audio and VoiceChat's 448ms turn-taking are not the same race: one synthesizes audio from supplied text, the other judges turns inside a bidirectional stream. v4 Turbo's 100ms and 150ms exclude network and application overhead; real-world feel adds model and pipeline latency, and third-party measured first audio near 165ms is a fairer reference (second-hand). Vendor figures obey one rule: relative, never absolute.
Closing: GPT-3 moment, or marketing moment
Six products in, the verdict is undramatic. The route war between true duplex and text mediation has just gone public; the open camp's consumer option is missing; official latency numbers uniformly exclude the network; and the likeliest next-generation model has entered no public benchmark. Dessa's researchers call it "the GPT-3 moment for speech." Whether that lands depends on weights, an API, or third-party evaluation arriving. Until then, sorting by route and tagging numbers by source is the reader's only reliable defense.
References
- Block's official Dessa page: architecture and local-run claims (official)
- TechCrunch, Oct 24, 2026: no public access, researcher-only early access (second-hand)
- Fish Audio Drama 3 blog: architecture, license, latency (official)
- ElevenLabs v4 and v4 Turbo agents page: latency and pricing (official)
- Bailian Qwen-Audio-3.0-Realtime Flash docs: duplex control, turn modes, pricing (official docs)
- HuggingFace VoiceChat-11B model card: architecture, duplex metrics, 80GB floor (official, relayed)
- Volcano Engine release notes: the one official line on Seeduplex (official)
- Artificial Analysis scores (Big Bench Audio, dynamics, pronunciation), third-party relay (third-party)
Related reading
- Voice Cloning Five Ways: Which One Stays Local
- Tencent Hy-Nomas: An Interactive World Model, Open and Deep-Dived
- Ming-3.0 Open-Source Speech: A Deployment SOP
FAQ
Q1: How is a full-duplex voice model different from an ordinary voice assistant?
A1: Ordinary assistants run recognition, generation, and synthesis in series; one direction at a time. Full-duplex models listen and speak simultaneously -- interrupt anytime, and they back-channel and yield. The former processes text; the latter works on speech, keeping pauses, laughter, and half-finished words text cannot carry.
Q2: Do Drama 3 and Eleven v4 Turbo count as full-duplex?
A2: Not by their own descriptions. Drama 3 is officially a "directable TTS," text-mediated; v4 Turbo is text-mediated TTS over a bidirectional WebSocket, interruption handled at the product layer. Calling them full-duplex is a sorting error.
Q3: Can you run an open-source full-duplex model locally on consumer GPUs today?
A3: Only Moshi, from 2024: 7B, CC-BY 4.0, about 200ms on a 24GB card -- English and French only, no tool calling. VoiceChat 11B is newer yet demands 80GB VRAM and a research-only license. "Weights you cannot afford, runnable ones out of date" remains the state of open full-duplex.
Q4: Should official latency numbers be taken at face value?
A4: Check the fine print. v4 Turbo's 100ms and 150ms exclude application and network latency; Drama 3's 88.9ms is a single-stream TTS figure racing nothing against the 448ms duplex metric. Third-party readings -- about 165ms first audio for v4 Turbo -- land closer to lived experience (second-hand).
Q5: What is the most realistic path to Chinese-language realtime voice conversation right now?
A5: A closed API with domestic access and RMB pricing is the shortest route; among the six only Qwen-Audio-3.0-Realtime qualifies. Doubao's release notes confirm full-duplex end-to-end, but its key numbers are second-hand. Self-hosting means accepting Moshi's language limits or VoiceChat's hardware floor.