AssemblyAI retweeted
Today we're launching AssemblyAI-STT-3.7 - three new transcription models with industry-leading accuracy on voice agent calls, live conversations, and recorded audio.
We rewrote our entire inference engine for these models, so they're now even faster and more scalable.
All three models are live now:
AssemblyAI-STT-3.7-Realtime-Agent:
- Perfect for voice agents
- #1 in accuracy for noisy telephony voice agent calls
- 86ms to the final transcript, 2x faster than our prior model
- 32 languages
AssemblyAI-STT-3.7-Realtime:
- Perfect for live meetings and captions
- The lowest speaker diarization error rate on live calls and meetings
- 32 languages in one stream with automatic language detection
AssemblyAI-STT-3.7 for recorded audio:
- Leading accuracy across 32 languages
- The lowest speaker diarization error rate of any pre-recorded STT API
- A 5-minute file comes back in 9 seconds
All available today in our API.
Learn more here: assemblyai.com/blog/introduc…
AssemblyAI retweeted
Hosted my first every Hardware hackathon at @AssemblyAI
can't wait to see you all for the next one, sign up below!
AssemblyAI retweeted
@AssemblyAI Universal 3.6 Pro has the lowest WER on clean audio: 1.77%. Their final transcript lands 91ms after the caller stops
AssemblyAI retweeted
Have you shipped speech-to-speech to production?
2027 Voice Agent Architecture Survey, first 200 get $25: forms.gle/uC4qs9PwnHPX3Y1s9
Have you tried AssemblyAI-tts-1 on LiveKit Inference yet?
@AssemblyAI built its first TTS model for voice agents and trained it on real phone conversations. It gets to first audio in about 90ms, offers 4,000+ voices, supports 6 languages, and includes word timestamps, all for $15 per 1M characters.
It’s one more choice in the lineup, so give it a try and see how it sounds with your agent > docs.livekit.io/agents/model…
AssemblyAI CEO @YouveGotFox reveals how just 10 seconds of audio is enough to clone someone's voice with AI:
"We have over 4,000 voices available on our API, and we also support instant voice cloning. With as little as 10 seconds of someone’s voice, you can instantly clone it. Obviously, if you have the consent to do that."
"The voice for your AI, in a lot of ways, that is the user experience. I might want my voice to sound funny and sarcastic because that fits my brand."
"If you’re calling into a business, you want it to sound local. If you’re in a town in the UK, you want it to have the same regional accent as where you’re living."
@AssemblyAI
AssemblyAI retweeted
Hopping on @MTSlive with with @sophiadew at 1:30PM PT to talk about @AssemblyAI's TTS-1 model we shipped today!
UK ROGUE AGENT RESPONSE | QWEN NEURONS | TOUCHSCREEN MACBOOK www.teapawt.org/i/broadcasts/1dxYlaZzE…
AssemblyAI retweeted
AssemblyAI has launched its first Text to Speech model! AssemblyAI-TTS-1 enters the top 12 in our Provider Voice Arena and the top 10 in Controlled Voice
@AssemblyAI has launched its first Text to Speech model, AssemblyAI-TTS-1! We evaluated its performance across both Speech Arenas and our Pronunciation Robustness benchmark.
Key takeaways:
➤ Text to Speech, Provider Voice: AssemblyAI-TTS-1 scores an Elo of 1,227 across 1,626 arena appearances, ranking #12, just behind Breeze TTS 2 at 1,228 and ahead of Mistral's Voxtral 3 TTS at 1,221 and Inworld's Realtime TTS-2 Flash at 1,218. It ranks #12 in both English (US) and English (UK).
➤ Text to Speech, Controlled Voice: AssemblyAI-TTS-1 scores 1,110 across 1,674 appearances, ranking #10 with the same 8 cloned voices, behind Cartesia's Sonic 3.5 at 1,118 and ahead of Speechify's Simba 3.2 at 1,086. It ranks #6 in Knowledge Sharing.
➤ Pronunciation Robustness: AssemblyAI-TTS-1 scores 81.7% across 2,115 rated clips, ranking #16, behind HeyGen Voice at 83.2% and ahead of Inworld Realtime TTS-2 at 80.7%. It ranks #6 in Preserving Exact Sequences at 81.5%.
➤ Speed: AssemblyAI-TTS-1 generates 162 characters per second, faster than Cartesia's Sonic 3.6 at 148.6 and Mistral's Voxtral 3 TTS at 112.3.
➤ Pricing: AssemblyAI-TTS-1 is priced at $15/1M characters. Of the eleven models ranked above it on Provider Voice, only two cost less: Gemini 3.8 Flash-Lite TTS at $11/1M and Simba 3.2 at $6.6/1M. Eleven v4 is $80/1M and Sonic 3.6 $49/1M.
See more details and listen to samples below ⬇️
AssemblyAI retweeted
message for those who think latency matters more than voice models accuracy!
"The difference between the almost right word and the right word is really a large matter—'tis the difference between the lightning bug and the lightning.
~Mark Twain
AssemblyAI retweeted
models great at coding aren't always great at voice.
we are open sourcing Voice Standard Bench: 50+ LLMs evaluated over voice transcripts that matter!
what model you would use?
you can compare and try it yourself with LLMGateway on AssemblyAI
AssemblyAI retweeted
Replying to @AssemblyAI
@AssemblyAI Universal-3.6 Pro is now live on @bolna_dev!
It is fine-tuned on human agent conversations, and that shows in the details your agents act on. Entity-aware endpointing helps the agent avoid cutting off a caller halfway through an account number.
Accuracy has improved, and it now supports 32 languages, including Hindi, Urdu, Marathi and more. Phone numbers, addresses, emails and codes are recognised and formatted according to how your agents need them.
PS: It’s free for all Bolna users from 9th to 16th of October. We’re excited to have you try it, so let us know how your agents perform.
AssemblyAI retweeted
Today at SF Tech Week, over 200 builders are reviving vintage toys with real voices at our hardware hackathon. 🧸🎙️
All turned into voice agents powered by @AssemblyAI 's speech models.
Can't wait to see what you all build. 8 hours to go !!
Replying to @AssemblyAI
@AssemblyAI new model, Universal 3.6 Pro is live on Vapi.
Most voice agent failures start at transcription. 3.6 Pro makes 45% fewer wrong yes/no confirmations, cuts background speech in transcripts by 30%, and holds the turn until it has the full phone number, code, or email. Plus 14 new languages with mid-sentence code-switching.
Switch your agents in the dashboard: dashboard.vapi.ai/
Impressive new streaming transcription model from @AssemblyAI.
We put a lot of work these days into creating good benchmarks and evals both for the individual models we use to build voice agents, and for complete voice agent implementations. Model launches like this are a good opportunity to talk about benchmarking!
The Assembly AI launch post cites two benchmarks: one that I work on and one from our friends at @covaldev. Both benchmarks show the new AssembleAI Universal 3.6 Pro Realtime model as having truly excellent accuracy and very good latency.
The accuracy of the new model is #1 in the Coval testing. It's #2 in our testing, edged out by the new Meta Muse Transcribe model. But the margin is so slim that there is unlikely to be any practical difference between the two in real-world use.
These two benchmarks use different data sets and different methodologies for calculating a transcription error rate score. Which is good, because diversity in testing is really helpful! You can read detailed descriptions of how error rates are calculated in the explainer pages for both benchmarks. And both benchmarks are fully open source, so you can read the code and reproduce the results yourself.
Both benchmarks also show the Universal 3.6 Pro Realtime model as having very good latency. For voice agent use cases, the important latency metric is the "time to final segment." This is how long it takes the model to deliver a complete transcript result after the user stops speaking.
The Coval benchmark shows a P95 TTFS of 225ms. We report 401ms. The Coval latency numbers are continously updated and our benchmark is a one-shot aggregate from 1,000 segments run all at once. So probably the Coval number is a better representation of what you'll see from the AssembleAI API.
In either case, these numbers put the new Assembly AI number on the Pareto frontier for all transcription models available today.
In plain English, this means that there's no model that's better on both accuracy and latency. There are models that are somewhat faster. And (one) model that's a tiny bit more accurate. But this new model is clearly a great choice for voice agents and moves the bar upwards in the increasingly competitive speech-to-text model category.
Universal 3.6 Pro Realtime is live today. Two independent benchmarks put every streaming speech model through the same audio, on the same harness. Universal-3.6 Pro Realtime leads both.
@trydaily's Pipecat open STT benchmark: Universal-3.6 Pro Realtime sits on the Pareto frontier.
@covaldev's live leaderboard: lowest word error rate of any streaming model on real voice-agent audio.
Voice agents get tested on the turns other benchmarks skip. That's where 3.6 Pro is built to hold:
→ Short answers. "No." "Nah." "Nuh-uh." Heard right 98.5% of the time, in loud rooms and on phone lines, not just quiet calls.
→ Background voices. The TV, the coworker, the next desk stay out of the transcript.Â
 → Languages. 32 from one endpoint with automatic detection. Callers switch languages mid-sentence and the transcript follows. Most agents can skip the language menu.
 → Turns. Entity-aware endpointing ends the turn when the caller is done, and holds it open while they're still reading out a phone number.
Trained on tens of thousands of hours of real voice-agent and telephony conversations. Link in the comments. 👇
AssemblyAI retweeted
Been out in the field a lot lately. Keep hearing: "With so many Voice AI vendors, what sets @AssemblyAI apart?"
Too many builders have been let down by the promises of models that look great in benchmarks and demos but fail in production.
Today that changes. AssemblyAI is launching Universal-3.6 Pro. THIS MODEL SLAPS 🔥
- Top performance on the hard stuff - entities that make or break voice workflows
- Agent context for biasing responses to agent questions
- General prompting for biasing specialized domains and audio conditions
- Entity-aware turn detection
- Robust performance in real-world conditions - noise, code-switching, specialized domains
Your users deserve an agent that actually hears them - and delivers an amazing experience.
Check out the research behind the model here: assemblyai.com/blog/universa…
AssemblyAI retweeted
AssemblyAI just launched Universal 3.6 Pro, and we ran it through our speech-to-text benchmark.
It now has the lowest word error rate of the 16 models we tested: 1.77% on 1,000 Pipecat clips, down from 1.93% on 3.5 Pro.
The bigger jump is final-text delay, the time from when a speaker stops to when the final transcript arrives. The median dropped from 180ms to 91ms, and P95 dropped from 692ms to 232ms.
Congrats to the @AssemblyAI team.
Full results: benchmarks.cekura.ai/stt
AssemblyAI retweeted
Your next app could start with your voice 🎙️
Build a dictation app with @AssemblyAI at Encode Hub this Thursday.
Get help from Voice AI engineers and credits to keep building.
📍 London · 1 Oct · 6-9pm BST
Bring your laptop. Leave with an app. Sign up now↓
AssemblyAI retweeted
AssemblyAI Universal 3.5 Pro topped our speech-to-text benchmark on both accuracy and latency.
We ran 15 models through the same audio and pipeline.
On 1,000 read-aloud clips from Pipecat FLEURS, Universal 3.5 Pro had the lowest word error rate at 1.93%.
It also had the fastest median time to first text at 489ms, so it started returning a transcript sooner than any other model we tested.
Congrats to the @AssemblyAI team.
Full results: benchmarks.cekura.ai/stt
AssemblyAI retweeted
Can’t wait to see what amazing voice apps you build in 15 mins at the workshop!
See you all in San Francisco: luma.com/xwnkujzr
The fastest way to get text into a computer is to talk.If your product has a text field, it has somewhere for dictation to go.
Thursday in SF, @iHarnoorSingh shows you how to put it there in about 30 lines of Python.
Save your seat: luma.com/xwnkujzr























