Zyphra Releases ZONOS3-Mini: Open-Source TTS With 100M Active Parameters

Step through</button><button class=”btn gh” id=”rst”>Reset</button></div>
</div>

<div class=”pn” id=”p2″>
<div class=”det”><strong>64 experts, top-2 routing.</strong> Each token activates 2 experts, so about 100M of the 800M parameters do work per step. That is why the model fits real-time budgets on laptops and phones.</div>
<div class=”moe” id=”moe”></div>
<div class=”cap” id=”moecap”>Routing token 1 …</div>
<div class=”ctl”><button class=”btn” id=”route”> Route tokens</button></div>
<div class=”note”>ZONOS2 used 8B parameters on an 86 Hz codec. ZONOS3-Mini cuts both: 800M total, 15 Hz frames.</div>
</div>

<div class=”pn” id=”p3″>
<div class=”grp”>Identity (speaker embedding + accent vectors)</div>
<div class=”chips” data-g=”id”>
<button class=”chip on” data-d=”Neutral base voice. No identity steering applied.”>Neutral</button>
<button class=”chip” data-d=”Age steering, e.g. age −1.5 targets a young adult and +1.5 an older adult.”>Age</button>
<button class=”chip” data-d=”Gender presentation steering, e.g. gender −1.5 for a masculine target.”>Gender</button>
<button class=”chip” data-d=”Accent task vectors, e.g. british 1.5 or australian 1.5, keeping the speaker recognizable.”>Accent</button>
<button class=”chip” data-d=”Texture steering: soft, raspy, bright or deep (e.g. raspy +5).”>Texture</button>
</div>
<div class=”grp”>Delivery (activation steering, word-level)</div>
<div class=”chips” data-g=”dl”>
<button class=”chip on” data-w=”-1″ data-d=”Whole line delivered neutrally.”>Neutral</button>
<button class=”chip” data-w=”0″ data-d=”Whisper +4 applied only to the span “I didn’t say”, via learned alignment heads.”>Whisper opening</button>
<button class=”chip” data-w=”1″ data-d=”Whisper +4 applied only to “she took the keys”.”>Whisper ending</button>
<button class=”chip” data-w=”-1″ data-d=”Emotion axes such as happy, angry or scared (e.g. scared +8).”>Emotion</button>
<button class=”chip” data-w=”-1″ data-d=”Pace −2 to +2 changes speed while keeping every word intelligible.”>Pace</button>
<button class=”chip” data-w=”-1″ data-d=”Projection −3 (intimate) to +3 (calling across a room).”>Projection</button>
</div>
<div class=”sent”><span class=”w” data-s=”0″>I</span> <span class=”w” data-s=”0″>didn’t</span> <span class=”w” data-s=”0″>say</span> <span class=”w” data-s=”1″>she</span> <span class=”w” data-s=”1″>took</span> <span class=”w” data-s=”1″>the</span> <span class=”w” data-s=”1″>keys.</span></div>
<div class=”det” id=”cdet”>Neutral base voice. No identity steering applied.</div>
<div class=”note”>Settings shown are examples from Zyphra’s published control demos. A VoiceCraft-style edit model can also change words after generation.</div>
</div>

<div class=”pn” id=”p4″>
<div class=”seg” id=”seg”>
<button class=”chip on” data-m=”ci”>Capability Index ↑</button>
<button class=”chip” data-m=”wer”>Clean WER ↓</button>
<button class=”chip” data-m=”sim”>Speaker similarity ↑</button>
<button class=”chip” data-m=”ut”>MSR-UTMOS ↑</button>
</div>
<div id=”bars”></div>
<div class=”note” id=”bnote”></div>
<div class=”note”>Source: Zyphra ZTTS2-Eval explorer, all-language set, ZONOS3-Mini in speaker-embedding mode. Self-reported by Zyphra.</div>
</div>

<div class=”pn” id=”p5″>
<div class=”grid” id=”dgrid”>
<div class=”dc on” data-c=”b”><b>Browser · WebGPU</b><p>Runs on your own GPU, no install. &lt;570 ms TTFA on a Pixel 9, &lt;150 ms on an M3 MacBook Air.</p></div>
<div class=”dc” data-c=”s”><b>NVIDIA · SGLang-Omni</b><p>OpenAI-compatible /v1/audio/speech with streaming. 80 ms TTFA on 1 H100 at 8 concurrent.</p></div>
<div class=”dc” data-c=”n”><b>Native · MLX / ggml</b><p>MLX for Apple Silicon; C++/ggml server and CLI for CPU, CUDA, Metal or Vulkan.</p></div>
<div class=”dc” data-c=”c”><b>Hosted · Zyphra Cloud</b><p>Managed API served on NVIDIA H200s.</p></div>
</div>
<pre id=”code”></pre>
</div>

<div class=”ft”><span>Sources: zyphra.com · huggingface.co/Zyphra · ZTTS2-Eval</span><span class=”mtp”>© Marktechpost</span></div>
</div>
<script>
(function(){
var R=document.getElementById(‘z3m’);
function $(s){return R.querySelector(s)} function $$(s){return R.querySelectorAll(s)}
function rs(){try{parent.postMessage({type:’z3m-h’,h:R.offsetHeight+40},’*’)}catch(e){}}
$$(‘.tab’).forEach(function(t){t.onclick=function(){$$(‘.tab’).forEach(function(x){x.classList.remove(‘on’)});$$(‘.pn’).forEach(function(x){x.classList.remove(‘on’)});t.classList.add(‘on’);$(‘#’+t.dataset.t).classList.add(‘on’);if(t.dataset.t===’p4′)draw(cur);rs()}});
var D=[‘<strong>Text.</strong> Input is NeMo-normalized (9 languages) and tokenized with an 8,192-entry SuperBPE vocabulary.’,'<strong>Speaker.</strong> Either a raw reference clip (up to 60 s, encoded into a token prefix) or a single 192-d MECT-B2 speaker embedding.’,'<strong>Hybrid MoE backbone.</strong> 11 layers: 9 Gated DeltaNet + 2 global attention, latent MoE with 64 experts and top-2 routing. Context: 32,768 positions.’,'<strong>ZAC-1 codes.</strong> Zyphra Audio Codec 1 at 15 Hz: 1 semantic + 9 acoustic codebooks of 4,096 entries each.’,'<strong>Causal iSTFT head.</strong> Each frame renders to 44.1 kHz audio as soon as it is complete, so streaming starts early.’];
var si=0,tm=null;
function setS(i){si=i;$$(‘#pipe .st’).forEach(function(x){x.classList.toggle(‘on’,+x.dataset.i===i)});$(‘#det’).innerHTML=D[i];}
$$(‘#pipe .st’).forEach(function(x){x.onclick=function(){setS(+x.dataset.i)}});
var F=$(‘#frames’);for(var k=0;k<30;k++){var d=document.createElement(‘div’);d.className=’fr’;F.appendChild(d)}
function frames(){var fs=F.children,j=0;var iv=setInterval(function(){if(j>=fs.length){clearInterval(iv);return}fs[j].classList.add(‘live’);fs[j].style.height=(25+Math.round(Math.abs(Math.sin(j*1.7))*75))+’%’;j++},70)}
$(‘#play’).onclick=function(){clearInterval(tm);setS(0);[].forEach.call(F.children,function(f){f.classList.remove(‘live’);f.style.height=’20%’});var n=0;tm=setInterval(function(){n++;if(n>4){clearInterval(tm);return}setS(n);if(n===4)frames()},1100)};
$(‘#rst’).onclick=function(){clearInterval(tm);setS(0);[].forEach.call(F.children,function(f){f.classList.remove(‘live’);f.style.height=’20%’})};
setS(0);
var M=$(‘#moe’);for(var e=0;e<64;e++){var x=document.createElement(‘div’);x.className=’ex’;M.appendChild(x)}
$(‘#route’).onclick=function(){var t=0;var iv=setInterval(function(){[].forEach.call(M.children,function(c){c.classList.remove(‘hit’)});if(t>=8){clearInterval(iv);$(‘#moecap’).textContent=’8 tokens routed · 2 of 64 experts active per token’;return}var a=Math.floor(Math.random()*64),b=(a+1+Math.floor(Math.random()*63))%64;M.children[a].classList.add(‘hit’);M.children[b].classList.add(‘hit’);t++;$(‘#moecap’).textContent=’Routing token ‘+t+’ → experts ‘+a+’ and ‘+b},450)};
$$(‘#p3 .chips’).forEach(function(g){g.querySelectorAll(‘.chip’).forEach(function(c){c.onclick=function(){g.querySelectorAll(‘.chip’).forEach(function(x){x.classList.remove(‘on’)});c.classList.add(‘on’);$(‘#cdet’).textContent=c.dataset.d;if(g.dataset.g===’dl’){var w=c.dataset.w;$$(‘#p3 .w’).forEach(function(s){s.classList.toggle(‘wh’,w!==’-1’&&s.dataset.s===w)})}}})});
var B={ci:{n:’Capability Index’,v:[66.2,61.7,60.2,37.7],max:70,note:’Item-response-theory index over 420 items. Reference-clip mode scores 44.7.’},wer:{n:’Clean WER’,v:[7.79,11.06,11.01,25.55],max:28,note:’Lower is better. Clean read speech, 35 languages.’},sim:{n:’Speaker similarity’,v:[83.4,81.2,82.9,69.6],max:90,note:’Higher is better. Clean set.’},ut:{n:’MSR-UTMOS’,v:[3.07,3.16,3.04,3.54],max:4,note:’Higher is better. ZONOS3-Mini trails Fish S2 Pro and Qwen3-TTS 1.7B here.’}};
var NM=[‘ZONOS3-Mini’,’Fish S2 Pro’,’MOSS-TTS v1.5′,’Qwen3-TTS 1.7B’],cur=’ci’;
function draw(m){cur=m;var b=B[m],h=”;for(var i=0;i<4;i++){h+='<div class=”row”><span class=”nm”>’+NM[i]+'</span><div class=”trk”><div class=”bar’+(i===0?’ me’:”)+'” data-w=”‘+(b.v[i]/b.max*100)+'”></div></div><span class=”val”>’+b.v[i]+'</span></div>’}$(‘#bars’).innerHTML=h;$(‘#bnote’).textContent=b.note;setTimeout(function(){$$(‘#bars .bar’).forEach(function(x){x.style.width=x.dataset.w+’%’})},40)}
$$(‘#seg .chip’).forEach(function(c){c.onclick=function(){$$(‘#seg .chip’).forEach(function(x){x.classList.remove(‘on’)});c.classList.add(‘on’);draw(c.dataset.m)}});
draw(‘ci’);
var C={b:’Open zyphra.com/zonos3mini/localplaygroundn# First visit downloads and caches the model.n# Needs a browser with WebGPU support.’,s:’sgl-omni serve –model-path Zyphra/ZONOS3-Mini –port 8000nncurl -X POST http://localhost:8000/v1/audio/speech \n -H “Content-Type: application/json” \n -d ‘{“input”: “Hello from ZONOS3-Mini!”}’ –output output.wav’,n:’# github.com/Zyphra/ZONOS3-Mininservers/mlx # Apple Siliconnservers/ggml # CPU, CUDA, Metal or Vulkan, no Python at runtimenservers/web # in-browser WebGPU build’,c:’cloud.zyphra.com/models/zyphra/ZONOS3-Minin# Hosted on NVIDIA H200s’};
function code(k){$(‘#code’).textContent=C[k];$$(‘#dgrid .dc’).forEach(function(x){x.classList.toggle(‘on’,x.dataset.c===k)});rs()}
$$(‘#dgrid .dc’).forEach(function(x){x.onclick=function(){code(x.dataset.c)}});code(‘b’);
window.addEventListener(‘load’,rs);window.addEventListener(‘resize’,rs);setTimeout(rs,300);
})();
</script>
</body></html>
“>