Prime Intellect has released Prime Mega MoE, a training extension of DeepSeek’s fused Mega MoE kernel. DeepSeek’s kernel only ran the forward pass. Prime Intellect added a fast, overlapping backward pass. That makes the fused kernel usable for training, not just inference. In its prime-rl trainer, full training steps on Qwen3 MoE models run up to 34% faster. Against TorchTitan, the gain reaches 55%. For teams training mixture-of-experts (MoE) models, that is a direct cut in GPU hours. TL;DR Size: A GPU kernel, not a model. Tested on Qwen3-30B-A3B and Qwen3-235B-A22B (128 experts, top-8 routing). Runs on: NVIDIA Hopper (blockwise FP8) and Blackwell (MXFP8) GPUs, BF16 on both. Published numbers use 8×H200 nodes, 32 to 128 GPUs. Performance: 32% to 34% faster full training steps than the prime-rl baseline. 48% to 55% faster than TorchTitan. Best: 3.11x faster forward pass on a Qwen3-30B-A3B MoE layer (12.0 ms to 3.9 ms). Worst: 1.24x backward pass on a Qwen3-235B-A22B MoE layer (30.5 ms to 24.6 ms). What is Prime Mega MoE? Prime Mega MoE is a fused, overlapping MoE kernel for training from Prime Intellect. It builds on the Mega MoE kernel in DeepSeek’s open-source DeepGEMM library. Why is MoE training communication-bound? In an MoE model, each token goes to a small subset of experts. Those experts sit on different GPUs. So every layer sends tokens out, runs expert math, and sends results back. This happens in the forward pass and again in the backward pass. In a standard setup, these are separate operations. The GPU sorts tokens and waits on the transport layer. Then it computes, hands results back, and waits again. Those waits are idle GPU time. How does the fused kernel work? DeepSeek’s Mega MoE, added to DeepGEMM in April 2026, replaces that chain with one kernel call. It fuses expert-parallel dispatch, the 2 expert linear layers, SwiGLU, and combine. While one token group is in transit, the kernel computes on a group that already arrived. It also skips intermediate tensors by using GPU-only scratch buffers. But it only shipped a forward implementation. Prime Intellect wrote the missing backward kernel, also overlapping. Supported types are BF16, MXFP8 (Blackwell only) and blockwise FP8 (Hopper only). The prime-rl config also exposes an fp8_bwd mode: BF16 forward, FP8 backward. (function(){var f=document.getElementById("mtp-pmm-frame");window.addEventListener("message",function(e){if(e.data&&e.data.mtpEmbed==="pmm"&&e.data.height&&f&&e.source===f.contentWindow){f.style.height=e.data.height+"px";}});})(); How fast is Prime Mega MoE? All numbers below are self-reported by Prime Intellect. Training steps use 8×H200 nodes, FP8, 16,384-token sequences and full activation checkpointing. Each result is the mean of steps 6 to 20. Full training step, seconds (lower is better): Against TorchTitan, throughput rose from 5,337 to 7,886 tokens/s/GPU on the 30B model. On the 235B model, it rose from 1,094 to 1,697. The prime-rl baseline uses torch all-to-all plus DeepGEMM FP8 experts. Single MoE layer, milliseconds (1 node, 8×H200, EP=8, 128 experts, top-8, random routing): The pattern is clear. Forward gains are large, and backward gains are smaller. The headline “up to 3.11x” refers to the forward pass only. How does it compare with other fused MoE options? TorchTitan’s 41% gain comes from a different setup (B200, 671B model, BF16 baseline). It is not directly comparable. How do developers use it? Prime Mega MoE is a dispatch backend in prime-rl. Set type = "mega_moe" under [trainer.model.moe.dispatch]. Choose precision from bf16, mxfp8, fp8 or fp8_bwd. The training kernels ship in a prebuilt DeepGEMM wheel from prime-kernels. The stock DeepGEMM build lacks them. Key Takeaways Prime Mega MoE adds a fused, overlapping backward pass to DeepSeek’s Mega MoE. Full training steps run 32% to 34% faster than the prime-rl baseline. Against TorchTitan, steps run 48% to 55% faster on Qwen3 MoE models. Forward gains reach 3.11x, but backward gains drop to 1.24x on large layers. It ships in prime-rl as the mega_moe dispatch backend. Check out the Prime Mega MoE blog post, the announcement on X, and the prime-rl repo. Also, feel free to follow us on Twitter and don’t forget to join our 150k+ML SubReddit and Subscribe to our Newsletter. Wait! are you on telegram? now you can join us on telegram as well. Need to partner with us for promoting your GitHub Repo OR Hugging Face Page OR Product Release OR Webinar etc.? Connect with us The post Prime Intellect Releases Prime Mega MoE: Trainable Fused MoE Kernel, 3.11x Faster Forward Pass appeared first on MarkTechPost.
Mistral has released the weights for Large 4, aka "Le Chonk," a trillion-parameter model trained entirely in Europe. Mistral says it beats its open rivals at coding, and independent testing shows it's fast and fairly cheap. It still can't match China's top models, scoring 26 points to the 32 posted by DeepSeek V4.1 Flash and GLM-5.3. The article Paris-based Mistral releases "Le Chonk," a strong coding model that still trails China overall appeared first on The Decoder.
The US Justice Department is looking into Nvidia's planned acquisition of open-source platform Hugging Face. The article US antitrust regulators are scrutinizing Nvidia's Hugging Face deal and sales practices appeared first on The Decoder.
US intelligence agencies believe an executive in Nvidia CEO Jensen Huang's circle is an informant for China. The article Chinese spies targeted Nvidia CEO Jensen Huang, according to The New York Times appeared first on The Decoder.
Chinese AI labs are now mainly tapping OpenAI's most powerful model, GPT-6 Astra, instead of Anthropic's Claude. The article OpenAI's Astra becomes the main target for Chinese AI labs after Anthropic tightens its defenses appeared first on The Decoder.
Google's stock dropped as much as ten percent after its quarterly earnings. Cloud revenue more than doubled to $30.7 billion, but search revenue grew just 12 percent, slower than expected. Google acknowledged that its fast-growing AI mode brings in less revenue than traditional search, and it is now expanding its AI mode ad program internationally. The article Google's AI search problem finally shows up in the earnings report, some investors believe appeared first on The Decoder.
Block’s AI research lab has released Dessa, a full-duplex speech model trained only on human conversation. Block announced it on October 23, 2026 as a research preview. Dessa is a speech-to-speech model from Block that listens and talks at the same time. Block says it used no text and no pretrained models. That is a different design from the text-first stacks behind most voice agents today. TL;DR Runs on: A MacBook Pro M4 Max or better, locally, faster than real time. Performance: Demo-only so far. Live recordings show backchannels, mid-sentence laughs and natural interruptions. Best: Trained on 1 NVIDIA GB200 NVL72 rack (72 GPUs). Moshi used 1,016 H100 GPUs. Worst: 0 published benchmark scores. No weights, paper, API or price. What is Block’s Dessa? Dessa is a true speech-to-speech model: audio in, audio out. According to Block’s project page, everything it knows came from human speech. It is full-duplex, so it listens while it talks. Block says it learned from real conversation when to cut in. It also learned when to backchannel with short cues like “yeah” or “got it”. Block’s demo clips show Dessa laughing mid-sentence and getting louder when asked to speak up. One clip shows it talking with OpenAI’s GPT-Live-1. Block says nobody steered any clip. How does speech-only training work? Most voice systems chain 3 parts: speech recognition, a text model, then speech synthesis. Block team argues that chain throws away emotion, sarcasm, laughter and sighs. Dessa hears and produces those signals directly. Block has not published the architecture. Here is what it has disclosed: Training data: human conversation audio only. No text and no pretrained models. Compute: 1 NVIDIA GB200 NVL72, which links 72 Blackwell GPUs and 36 Grace CPUs. Data efficiency: Block says Dessa learned language from hundreds of times fewer words than LLMs need. Inference: faster than real time, on a MacBook Pro M4 Max or better. The lab also explains why this is hard. Speech never repeats exactly. A spoken word takes thousands of times more bytes than the same word in text. Kyutai’s Moshi, the best-known open full-duplex model, takes another route. Moshi starts from a text language model backbone and predicts time-aligned text tokens before audio tokens. Its model card lists 7 million hours of pretraining audio and 1,016 H100 GPUs. (function(){var f=document.getElementById('mtp-dessa-frame'); window.addEventListener('message',function(e){if(e.source!==f.contentWindow)return;var d=e.data;if(d&&d.mtpDessaH){f.style.height=d.mtpDessaH+'px';}});})(); How does Dessa compare with Moshi, PersonaPlex and GPT-Live-1? Key Takeaways Dessa is full-duplex and speech-to-speech, trained only on human conversation audio. Block trained it on 1 GB200 NVL72 rack; Moshi used 1,016 H100s. It runs locally on a MacBook Pro M4 Max, faster than real time. Block published 0 benchmarks, no parameter count and no weights. Access is a waitlist. Check out the Dessa project page. Compare it with the Moshi paper, the PersonaPlex model card and the GPT-Live-1 docs. Also, feel free to follow us on Twitter and don’t forget to join our 150k+ML SubReddit and Subscribe to our Newsletter. Wait! are you on telegram? now you can join us on telegram as well. Need to partner with us for promoting your GitHub Repo OR Hugging Face Page OR Product Release OR Webinar etc.? Connect with us The post Block Unveils Dessa, a Full-Duplex Speech Model Trained Only on Speech appeared first on MarkTechPost.
AssemblyAI has released AssemblyAI-TTS-1, its first text-to-speech model, now generally available. The AssemblyAI-TTS-1 text-to-speech API targets voice agents running on live phone calls. In that setting, latency and pronunciation decide whether a caller stays on the line. Pricing is $15 per 1M characters, with no annual contract and no concurrency tier to buy. For developers already using AssemblyAI speech-to-text, it closes the voice loop with one vendor. TL;DR Size: It is a hosted API model. Runs on: AssemblyAI’s cloud only, in the US (us-east-1) or EU (eu-north-1). No open weights or on-prem option. Any client with HTTP or WebSocket works. Performance: Ties ElevenLabs v4 on entity accuracy in AssemblyAI’s tests, at under 1/5 of its price. Best: 47 ms median time to first byte across 6 models in AssemblyAI’s API test (self-reported). Worst: 159 ms median streaming latency on Cekura’s independent benchmark, 8th of 19 timed models. Bottom line: Best: $15 per 1M characters with unlimited concurrency. Worst: no SSML, speed control, pronunciation dictionaries or emotion tags yet. What is AssemblyAI-TTS-1? AssemblyAI-TTS-1 is a streaming text-to-speech model from AssemblyAI that turns agent replies into speech for live calls. It was trained on millions of hours of audio, including real conversational data. Most TTS models represent audio as discrete tokens. AssemblyAI says TTS-1 operates on continuous audio representations instead. The AssemblyAI claims this improves realism and cuts time to first spoken audio. The launch includes a library of 4,000+ voices and instant voice cloning. A clone needs a short recording and a confirmation of the speaker’s permission. The clone in the launch post used 24 seconds of audio. Six languages are supported at launch: English, Spanish, French, German, Italian and Portuguese. How do developers integrate AssemblyAI-TTS-1? There are 3 interfaces: an HTTP API, a WebSocket API and an OpenAI-compatible route. The WebSocket session keeps one connection per conversation and accepts a cancel message for barge-in. Output is mono PCM, mu-law or A-law, at 8 kHz to 48 kHz. 24 kHz PCM is the native format; telephony stacks can request 8 kHz mu-law directly. Word timestamps are included. AssemblyAI also published LiveKit and Pipecat reference agents. How does AssemblyAI-TTS-1 perform on benchmarks? AssemblyAI tested TTS-1 against 5 models on 4 English sets, using 50 matched cloned voices. All figures below are self-reported, from an eval dated October 28, 2026. Entity accuracy (580 hard targets such as phone numbers and codes): 93.7%, vs ElevenLabs v4 93.6%, Cartesia Sonic 3.6 82.5%, ElevenLabs Flash v2.5 68.0%. Word accuracy: 99.0% on Seed-TTS and LibriSpeech sentences. Predicted MOS: 3.85, behind Inworld TTS-2 Flash at 3.88 and ahead of Sonic 3.6 at 3.82. Median time to first byte: 47 ms. Flash v2.5 took 119 ms, Sonic 3.6 143 ms, Inworld TTS-2 Flash 244 ms. Other: 85.5 TTSDS prosody score and 0.64 speaker similarity for cloned audio. Independent results are more mixed. On Cekura’s benchmark of 27 models, TTS-1 scored 81.5% reading accuracy and 4.50 naturalness. Its median streaming latency, timed from the last word sent, was 159 ms. Inworld TTS-2 Flash posted 45 ms and 93.2% accuracy there, at the same $15 price. AssemblyAI also cites a 1.7% word error rate on Coval’s leaderboard. It reports a top-10 debut in Artificial Analysis’ controlled-voice arena. <iframe id="mtp-aai-tts1-frame" title="AssemblyAI-TTS-1 interactive explainer" loading="lazy" style="width:100%;height:600px;border:0;display:block;background:transparent;" scrolling="no" srcdoc="<!doctype html><title>AssemblyAI-TTS-1 Explainer</title>html,body{margin:0;padding:0;background:transparent} #mtp-aai-tts1{--cob:#3923c7;--cob3:#887bdd;--cob2:#b0a7e9;--ink:#1d1b16;--ink2:#2a2721;--line:#4a4945;--cream:#fdfcf8;--muted:#a5a4a2;--blush:#e39389;--mtp:#76B900; max-width:880px!important;margin:0 auto!important;background:var(--ink)!important;color:var(--cream)!important;border:1px solid #3a3730!important;border-radius:16px!important;font-family:Inter,system-ui,-apple-system,Segoe UI,Roboto,sans-serif!important;overflow:hidden;box-sizing:border-box;line-height:1.5} #mtp-aai-tts1 *{box-sizing:border-box} #mtp-aai-tts1 hr,#mtp-aai-tts1 p:empty,#mtp-aai-tts1 del,#mtp-aai-tts1 s{display:none!important} #mtp-aai-tts1 .hd{padding:22px 26px 14px;background:radial-gradient(ellipse at 90% -20%,rgba(57,35,199,.55),transparent 60%)!important} #mtp-aai-tts1 .eyebrow{font:600 11px/1 ui-monospace,Menlo,monospace;letter-spacing:.14em;text-transform:uppercase;color:var(--cob2)!important} #mtp-aai-tts1 h3{margin:8px 0 4px!important;font-size:24px!important;font-weight:700!important;color:var(--cream)!important;border:0!important;padding:0!important} #mtp-aai-tts1 .sub{color:var(--muted)!important;font-size:14px;margin:0!important} #mtp-aai-tts1 .tabs{display:flex;gap:6px;padding:0 26px;border-bottom:1px solid #3a3730!important;overflow-x:auto} #mtp-aai-tts1 .tab{background:transparent!important;border:0!important;border-bottom:2px solid transparent!important;color:var(--muted)!important;font:600 13px Inter,system-ui,sans-serif;padding:12px 10px;cursor:pointer;white-space:nowrap;transition:color .2s} #mtp-aai-tts1 .tab.on{color:var(--cream)!important;border-bottom-color:var(--cob3)!important} #mtp-aai-tts1 .pane{display:none;padding:22px 26px 10px;animation:mtpfade .35s ease} #mtp-aai-tts1 .pane.on{display:block} @keyframes mtpfade{from{opacity:0;transform:translateY(6px)}to{opacity:1;transform:none}} #mtp-aai-tts1 .flow{display:flex;align-items:stretch;gap:8px;margin:4px 0 14px} #mtp-aai-tts1 .node{flex:1;background:var(--ink2)!important;border:1px solid #3a3730!important;border-radius:12px;padding:12px 10px;text-align:center;font-size:12.5px;color:var(--muted)!important;transition:all .35s;position:relative} #mtp-aai-tts1 .node b{display:block;color:var(--cream)!important;font-size:13.5px;margin-bottom:3px} #mtp-aai-tts1 .node.act{border-color:var(--cob3)!important;background:rgba(57,35,199,.28)!important;color:var(--cob2)!important;transform:translateY(-3px);box-shadow:0 8px 24px rgba(57,35,199,.35)} #mtp-aai-tts1 .arrow{align-self:center;color:#4a4945!important;font-size:18px} #mtp-aai-tts1 .wave{display:flex;align-items:center;gap:3px;height:34px;justify-content:center;margin-top:6px} #mtp-aai-tts1 .wave i{display:block;width:4px;border-radius:2px;background:var(--cob3)!important;height:6px;opacity:.35} #mtp-aai-tts1 .playing .wave i{animation:mtpwave 1s ease-in-out infinite;opacity:1} @keyframes mtpwave{0%,100%{height:6px}50%{height:30px}} #mtp-aai-tts1 .detail{background:#24211b!important;border-left:3px solid var(--cob3)!important;border-radius:8px;padding:12px 14px;font-size:14px;min-height:66px;color:var(--cream)!important} #mtp-aai-tts1 .ctrl{display:flex;gap:8px;margin:12px 0 4px;flex-wrap:wrap} #mtp-aai-tts1 .btn{background:var(--cob)!important;color:#fff!important;border:0!important;border-radius:999px;padding:8px 16px;font:600 13px Inter,system-ui,sans-serif;cursor:pointer;transition:background .2s} #mtp-aai-tts1 .btn:hover{background:#614fd2!important} #mtp-aai-tts1 .btn.ghost{background:transparent!important;border:1px solid #4a4945!important;color:var(--cream)!important} #mtp-aai-tts1 .pills{display:flex;gap:6px;flex-wrap:wrap;margin-bottom:14px} #mtp-aai-tts1 .pill{background:transparent!important;border:1px solid #4a4945!important;color:var(--muted)!important;border-radius:999px;padding:6px 12px;font:600 12px Inter,system-ui,sans-serif;cursor:pointer} #mtp-aai-tts1 .pill.on{background:var(--cob)!important;border-color:var(--cob)!important;color:#fff!important} #mtp-aai-tts1 .bars{display:flex;flex-direction:column;gap:9px} #mtp-aai-tts1 .row{display:grid;grid-template-columns:150px 1fr 70px;gap:10px;align-items:center;font-size:13px} #mtp-aai-tts1 .row .nm{color:var(--cream)!important} #mtp-aai-tts1 .track{display:block;height:14px;background:#2a2721!important;border-radius:7px;overflow:hidden} #mtp-aai-tts1 .fill{display:block;height:100%;width:0;border-radius:7px;background:#5a574f!important;transition:width .9s cubic-bezier(.2,.8,.2,1)} #mtp-aai-tts1 .row.me .fill{background:linear-gradient(90deg,#3923c7,#887bdd)!important} #mtp-aai-tts1 .row.me .nm{color:var(--cob2)!important;font-weight:700} #mtp-aai-tts1 .val{text-align:right;font-variant-numeric:tabular-nums;color:var(--cream)!important} #mtp-aai-tts1 .note{font-size:12px;color:var(--muted)!important;margin:12px 0 0!important} #mtp-aai-tts1 .calc label{font-size:13px;color:var(--muted)!important} #mtp-aai-tts1 input[type=range]{width:100%;accent-color:#887bdd;margin:8px 0 4px} #mtp-aai-tts1 .big{font-size:26px;font-weight:700;color:var(--cream)!important} #mtp-aai-tts1 .grid{display:grid;grid-template-columns:1fr 1fr;gap:10px;margin-top:14px;align-items:start} #mtp-aai-tts1 .card{background:var(--ink2)!important;border:1px solid #3a3730!important;border-radius:12px;padding:12px 14px;cursor:pointer;transition:border-color .2s} #mtp-aai-tts1 .card:hover{border-color:var(--cob3)!important} #mtp-aai-tts1 .card h4{margin:0!important;font-size:14px!important;color:var(--cream)!important;display:flex;justify-content:space-between} #mtp-aai-tts1 .card p{margin:8px 0 0!important;font-size:13px;color:var(--muted)!important;display:none} #mtp-aai-tts1 .card.open p{display:block;animation:mtpfade .3s ease} #mtp-aai-tts1 .card .tag{font:600 10px ui-monospace,monospace;color:var(--blush)!important;letter-spacing:.08em} #mtp-aai-tts1 .cost{display:grid;grid-template-columns:1fr auto;gap:6px 12px;font-size:14px;margin-top:10px} #mtp-aai-tts1 .cost .me{color:var(--cob2)!important;font-weight:700} #mtp-aai-tts1 .ft{display:flex;justify-content:space-between;gap:10px;flex-wrap:wrap;padding:14px 26px 18px;font-size:11.5px;color:var(--muted)!important;border-top:1px solid #3a3730!important;margin-top:14px} #mtp-aai-tts1 .ft .mtp{color:var(--mtp)!important;font-weight:700} #mtp-aai-tts1 a{color:var(--cob2)!important} @media (max-width:640px){ #mtp-aai-tts1 .hd,#mtp-aai-tts1 .pane{padding-left:16px;padding-right:16px} #mtp-aai-tts1 .tabs{padding:0 12px} #mtp-aai-tts1 .flow{flex-direction:column} #mtp-aai-tts1 .arrow{transform:rotate(90deg)} #mtp-aai-tts1 .row{grid-template-columns:110px 1fr 58px;font-size:12px} #mtp-aai-tts1 .grid{grid-template-columns:1fr} #mtp-aai-tts1 h3{font-size:20px!important} } <div id="mtp-aai-tts1"> <div class="hd"> <div class="eyebrow">AssemblyAI · Text-to-Speech · Oct 2026</div> <h3>AssemblyAI-TTS-1, explained</h3> <p class="sub">A streaming voice for phone agents: $15 per 1M characters, 4,000+ voices, 6 languages.</p> </div> <div class="tabs" role="tablist"> <button class="tab on" data-t="flow">How it works</button> <button class="tab" data-t="bench">Benchmarks</button> <button class="tab" data-t="cost">Cost calculator</button> <button class="tab" data-t="limits">Limits</button> </div> <div class="pane on" data-p="flow"> <div class="flow" id="aai-flow"> <div class="node" data-i="0"><b>Agent LLM</b>writes the reply</div><div class="arrow">›</div> <div class="node" data-i="1"><b>WebSocket</b>1 session per call</div><div class="arrow">›</div> <div class="node" data-i="2"><b>TTS-1</b>continuous audio</div><div class="arrow">›</div> <div class="node" data-i="3"><b>Audio out</b>8 to 48 kHz<div class="wave"><i></i><i></i><i></i><i></i><i></i><i></i><i></i></div></div><div class="arrow">›</div> <div class="node" data-i="4"><b>Barge-in</b>cancel message</div> </div> <div class="detail" id="aai-detail">Press Play or Next to step through a live call turn.</div> <div class="ctrl"> <button class="btn" id="aai-play"> Play</button> <button class="btn ghost" id="aai-prev">‹ Back</button> <button class="btn ghost" id="aai-next">Next ›</button> </div> </div> <div class="pane" data-p="bench"> <div class="pills" id="aai-pills"> <button class="pill on" data-m="ent">Entity accuracy</button> <button class="pill" data-m="ttfb">Median TTFB (AssemblyAI test)</button> <button class="pill" data-m="cstream">Streaming latency (Cekura)</button> <button class="pill" data-m="cacc">Reading accuracy (Cekura)</button> </div> <div class="bars" id="aai-bars"></div> <p class="note" id="aai-bnote"></p> </div> <div class="pane" data-p="cost"> <div class="calc"> <label for="aai-min">Spoken minutes per month: <b id="aai-minv" style="color:#fdfcf8">10,000</b></label> <div class="cost" id="aai-cost"></div> <p class="note">Assumes 600 characters per minute of phone speech, the figure AssemblyAI cites as typical. List prices per 1M characters: AssemblyAI-TTS-1 $15, Inworld TTS-2 Flash $15 on demand, ElevenLabs Flash v2.5 $40, ElevenLabs v4 $80. Concurrency limits are not priced in.</p> </div> </div> <div class="pane" data-p="limits"> <div class="grid" id="aai-cards"> <div class="card"><h4>No SSML or speed control <span class="tag">LAUNCH GAP</span></h4><p>The WebSocket API has no speed, pitch, emotion or style parameter, and neither API accepts SSML. Punctuation and written fillers like "um" still shape delivery.</p></div> <div class="card"><h4>No pronunciation dictionaries <span class="tag">COMING SOON</span></h4><p>Custom pronunciation dictionaries and emotion tags are announced but not shipped. Cartesia Sonic 3.6 already offers custom pronunciations.</p></div> <div class="card"><h4>6 languages <span class="tag">COVERAGE</span></h4><p>English, Spanish, French, German, Italian and Portuguese. ElevenLabs Flash v2.5 lists 32 and Cartesia Sonic 3.6 lists 44.</p></div> <div class="card"><h4>Latency depends on method <span class="tag">CHECK IT</span></h4><p>About 90 ms advertised, 47 ms median in AssemblyAI's own test, 159 ms median streaming latency on Cekura (timed from the last word sent).</p></div> <div class="card"><h4>Instant cloning only <span class="tag">LAUNCH GAP</span></h4><p>Pro (fine-tuned) voice cloning is not available yet. Instant clones need a short clip and speaker permission.</p></div> <div class="card"><h4>Cloud only <span class="tag">DEPLOYMENT</span></h4><p>US (us-east-1) and EU (eu-north-1) hosts at the same price. No on-prem option is announced.</p></div> </div> <p class="note">Tap a card to expand it.</p> </div> <div class="ft"><span>Sources: AssemblyAI launch post, docs and pricing; Cekura TTS benchmark; vendor pricing pages. Checked Oct 29, 2026.</span><span class="mtp">Built by Marktechpost</span></div> </div> (function(){ var R=document.getElementById('mtp-aai-tts1');if(!R)return; function resize(){try{parent.postMessage({mtpEmbed:'aai-tts1',h:R.offsetHeight+40},'*')}catch(e){}} R.querySelectorAll('.tab').forEach(function(t){t.addEventListener('click',function(){ R.querySelectorAll('.tab').forEach(function(x){x.classList.remove('on')});t.classList.add('on'); R.querySelectorAll('.pane').forEach(function(p){p.classList.toggle('on',p.getAttribute('data-p')===t.getAttribute('data-t'))}); if(t.getAttribute('data-t')==='bench')draw(cur);setTimeout(resize,60);});}); var steps=['Your agent LLM writes a reply. Text can be sent word by word as it is generated.', 'One WebSocket session per conversation. Voice, language, sample rate and encoding are fixed at connect time.', 'TTS-1 models speech on continuous audio representations, not discrete tokens. AssemblyAI says this improves realism and speed.', 'Audio streams back as raw mono PCM, mu-law or A-law. 24 kHz PCM is native; 8 kHz mu-law suits phone lines.', 'If the caller interrupts, a cancel message stops speech. Word timestamps show what was actually said.']; var nodes=R.querySelectorAll('#aai-flow .node'),det=R.querySelector('#aai-detail'),si=-1,timer=null; function show(i){si=(i+nodes.length)%nodes.length;nodes.forEach(function(n,k){n.classList.toggle('act',k===si);n.classList.toggle('playing',k===3&&si>=3)});det.textContent=(si+1)+'/5 · '+steps[si];resize();} R.querySelector('#aai-next').onclick=function(){stop();show(si+1)}; R.querySelector('#aai-prev').onclick=function(){stop();show(si<=0?4:si-1)}; var pb=R.querySelector('#aai-play'); function stop(){if(timer){clearInterval(timer);timer=null;pb.textContent=' Play'}} pb.onclick=function(){if(timer){stop();return}pb.textContent='❚❚ Pause';show(si+1);timer=setInterval(function(){show(si+1)},2200)}; var D={ ent:{u:'%',max:100,hi:true,note:'AssemblyAI pronunciation set: 580 hard targets (phone numbers, dates, URLs, codes). Self-reported, eval dated Oct 28, 2026.', r:[['AssemblyAI-TTS-1',93.7,1],['ElevenLabs v4',93.6],['Cartesia Sonic 3.6',82.5],['ElevenLabs Flash v2.5',68.0]]}, ttfb:{u:' ms',max:650,hi:false,note:'Median time to first byte over each API from one client, one request at a time. Self-reported by AssemblyAI. Lower is better.', r:[['AssemblyAI-TTS-1',47,1],['ElevenLabs Flash v2.5',119],['Cartesia Sonic 3.6',143],['Inworld TTS-2 Flash',244],['Inworld TTS-2',280],['ElevenLabs v4',612]]}, cstream:{u:' ms',max:600,hi:false,note:'Cekura: median from the last word sent to first audible sound, US East, 104 sentences x 3 reads. Independent. Lower is better.', r:[['Inworld TTS-2 Flash',45],['AssemblyAI-TTS-1',159,1],['ElevenLabs Flash v2.5',177],['Cartesia Sonic 3.6',270],['ElevenLabs v4',561]]}, cacc:{u:'%',max:100,hi:true,note:'Cekura: hard parts read correctly, scored by speech-to-text transcribers (an estimate, not human-rated). Independent.', r:[['Inworld TTS-2 Flash',93.2],['AssemblyAI-TTS-1',81.5,1],['ElevenLabs v4',80.4],['Cartesia Sonic 3.6',74.9],['ElevenLabs Flash v2.5',64.4]]}}; var cur='ent',bars=R.querySelector('#aai-bars'),bn=R.querySelector('#aai-bnote'); function draw(m){cur=m;var d=D[m];bars.innerHTML='';d.r.forEach(function(x){ var row=document.createElement('div');row.className='row'+(x[2]?' me':''); row.innerHTML='<span class="nm">'+x[0]+'</span><span class="track"><span class="fill"></span></span><span class="val">'+(d.u==='%'?x[1].toFixed(1):x[1])+d.u+'</span>'; bars.appendChild(row);var f=row.querySelector('.fill');setTimeout(function(){f.style.width=Math.max(2,x[1]/d.max*100)+'%'},40);}); bn.textContent=d.note;setTimeout(resize,60);} R.querySelectorAll('#aai-pills .pill').forEach(function(p){p.addEventListener('click',function(){ R.querySelectorAll('#aai-pills .pill').forEach(function(x){x.classList.remove('on')});p.classList.add('on');draw(p.getAttribute('data-m'))})}); var rng=R.querySelector('#aai-min'),mv=R.querySelector('#aai-minv'),co=R.querySelector('#aai-cost'); var P=[['AssemblyAI-TTS-1',15,1],['Inworld TTS-2 Flash (on demand)',15],['ElevenLabs Flash v2.5',40],['ElevenLabs v4',80]]; function calc(){var m=+rng.value,ch=m*600;mv.textContent=m.toLocaleString('en-US');var h=''; P.forEach(function(p){var c=ch/1e6*p[1];h+='<span class="'+(p[2]?'me':'')+'">'+p[0]+'</span><span class="'+(p[2]?'me':'')+'">$'+c.toLocaleString('en-US',{maximumFractionDigits:0})+'/mo</span>'}); co.innerHTML=h+'<span style="color:#a5a4a2">Characters per month</span><span style="color:#a5a4a2">'+(ch/1e6).toLocaleString('en-US',{maximumFractionDigits:1})+'M</span>';} rng.addEventListener('input',calc);calc(); R.querySelectorAll('#aai-cards .card').forEach(function(c){c.addEventListener('click',function(){c.classList.toggle('open');setTimeout(resize,40)})}); draw('ent');show(0); window.addEventListener('load',resize);window.addEventListener('resize',resize);setTimeout(resize,300); })(); "> (function(){var f=document.getElementById("mtp-aai-tts1-frame");window.addEventListener("message",function(e){if(e.data&&e.data.mtpEmbed==="aai-tts1"&&e.data.h){f.style.height=e.data.h+"px";}});})(); How does AssemblyAI-TTS-1 compare with ElevenLabs, Cartesia and Inworld? Cartesia’s range divides each plan price by its credits, at 1 credit per character. What does AssemblyAI-TTS-1 cost? Pricing is per input character, including spaces and punctuation, at $15 per 1M. Every voice, cloning, word timestamps and EU processing come at that same rate. There are no per-voice fees, concurrency fees or minimums. AssemblyAI calls it the lowest-priced frontier TTS model for voice agents. However, Inworld TTS-2 Flash also lists $15 per 1M on demand, falling to $7 on paid plans. The sharper difference is concurrency: Inworld’s on-demand tier allows 5 concurrent requests. Key Takeaways AssemblyAI-TTS-1 is AssemblyAI’s first text-to-speech model, built for live voice agents. It costs $15 per 1M characters, with unlimited concurrency and no contract. Self-reported: 93.7% entity accuracy and 47 ms median time to first byte. Cekura measured 159 ms streaming latency and 81.5% reading accuracy. Missing at launch: SSML, speed control, pronunciation dictionaries and emotion tags. Check out the launch post, docs, pricing, and the LiveKit and Pipecat reference agents. Also, feel free to follow us on Twitter and don’t forget to join our 150k+ML SubReddit and Subscribe to our Newsletter. Wait! are you on telegram? now you can join us on telegram as well. Need to partner with us for promoting your GitHub Repo OR Hugging Face Page OR Product Release OR Webinar etc.? Connect with us The post AssemblyAI Ships AssemblyAI-TTS-1: A $15 per 1M Character Text-to-Speech API for Voice Agents appeared first on MarkTechPost.
