alt.hn

8/17/2026 at 3:36:18 PM

Launch HN: Speko (YC S26) – OpenRouter for Voice AI

 https://speko.ai/

by abdik

8/19/2026 at 3:55:16 PM

Two things bit me when I was choosing a voice stack, and I don't see either as an axis in your benchmarks.

First: whether a model is "suitable for realtime" turned out to be a property of the transport, not of the model. I had written one TTS off as too slow based on the vendor's own guidance, then measured it again over a streaming path and got about 0.9s where the non-streaming call had taken 5s. Same model, same provider. If a benchmark is run over one transport, a model can look disqualified when it's actually fine for the way you'd ship it.

Second: non-English breakage doesn't show up in aggregate quality numbers. The speed-optimised tiers - the flash/turbo class - were fine in English and fell apart in Japanese. Not "slightly worse": confidently wrong words that changed what the sentence meant. What made that expensive is that the vendor's own docs said their turbo tier was equivalent to their flash tier. That's true in English and wasn't true in the language I was shipping in, so the documentation actively pointed me the wrong way.

Both of those mean the thing I'd actually pay a routing layer for is per-language and per-transport measurements, rather than one quality/latency/cost point per model. "Best TTS under 300ms" has a different answer in English than in Japanese, and I couldn't get that out of any vendor's published numbers.

So: are your benchmarks per-language, and do you measure over the transport people actually ship on? If yes, that's a bigger deal than the routing itself for anyone shipping outside English.

by noagogo

8/20/2026 at 3:24:32 AM

Routing makes even more sense for voice than text, but the constraint space is trickier: latency budget (round-trip vs streaming), WER on accented speech, and TTS naturalness all trade off non-linearly. In our voice pipeline, DeepSeek-V4-Pro plus a small dedicated STT beat an end-to-end frontier voice model on cost-per-minute by ~10x while staying inside a 300ms added-latency budget - purely because we could mix and match components. The hard part is honest benchmarking: WER numbers are only comparable within the same eval set, so a "find the optimal combo" service lives or dies by its methodology. Do you expose per-component benchmarks (STT WER, TTS MOS) separately, or only the combined scores?

by modgate

8/18/2026 at 10:24:47 PM

The piece I'd want to see in the constraint solver is effective cost rather than list price. On the LLM leg alone, current 5.6 pricing splits at 272K input tokens with the long band at exactly 2x, and cached input runs 90% below standard input, so two stacks with identical list prices can differ several-fold depending on how much of the prompt is a stable prefix.

Does the optimizer model cache-hit rate and context distribution, or does it score on list rates?

by hirak10

8/17/2026 at 5:05:54 PM

The benchmarks page seems interesting and something I can use to help make an informed decision. Can you talk about how you're measuring some of these? I imagine it needs to involve some human input.

https://benchmarks.speko.ai/turntaking

by webo

8/17/2026 at 7:19:03 PM

Turn-taking specifically does not need a listening panel, but it is measured mechanically. 200+ real human clips, and we score end-vs-wait decisions: did the model decide the caller finished speaking, or just paused mid-thought. The best detector gets 94.0% of those right; a plain VAD silence timer gets 46.9%. Results are published here: https://benchmarks.speko.ai/turntaking

You are right about human input for naturalness, that one we did not automate away with yet. We run blind A/B listening rounds with native speakers.

by abdik

8/17/2026 at 4:38:48 PM

Does this include a turn taking API? It'd be great to have one API that could do "Conversation in a box". One of the biggest annoyances is daisy chaining many models together for turn taking, dumb models for immediate responses, with smarter models returning and taking over after.

by spmartin823

8/17/2026 at 7:17:00 PM

Yes, on the hosted side (agents platform): full sessions come with VAD and turn-taking handled - we set them up and tune them for your use case, so that is the closest thing to conversation in a box. If you run your own orchestration, the gateway is just the routing layer and turn-taking stays in your framework - in our own stack we run Pipecat's Smart Turn in-process and tune the commit threshold on real calls. We also share our benchmarks here: https://benchmarks.speko.ai/turntaking

On fast dumb models answering while a smarter one takes over: we are experimenting with exactly that split - a small fast model holds the conversation while a larger one works behind it. Today it runs as two pinned routes, not one packaged API. Most turns in a phone call do not need a frontier model, and the fastest models on our LLM board are all small, so this is where routing earns its keep. We publish benchmarks on LLMs here: https://benchmarks.speko.ai/llm

by abdik

8/18/2026 at 3:56:42 PM

One packaged API for the takeover would be great, also worth considering predicting answers with a stronger LLM while the user is talking. Ello (the reading tutoring company) does this.

by spmartin823

8/18/2026 at 2:07:18 PM

Answering your question from the small end: I picked my STT by testing self-correction handling. My tool cleans up spoken drafts, and the failure that mattered wasn't word accuracy, it was "meet Tuesday, no wait, Wednesday": a raw transcript of that is worse than useless, and models differ a lot in how gracefully downstream cleanup can recover. Your spontaneous-speech testing sounds close to this already. Do the boards score corrections and disfluencies specifically, or do they fold into overall accuracy?

by jobuildsstuff

8/19/2026 at 9:01:58 AM

We hit the same problem building a voice-memo-to-notes tool: "meet Tuesday, no wait, Wednesday" is exactly the failure mode. We ended up not trying to fix it at the ASR layer at all. We run faster-whisper (large-v3, local) and let the raw transcript keep the self-correction intact, including the discarded "Tuesday." The summarization pass resolves it, because it has document-level context the transcriber doesn't, and it's cheap to re-run on a short transcript. A technically "wrong" transcript still produces the right structured output most of the time. The failure mode shifts rather than disappears though: on longer memos where corrections stack up, the summarizer occasionally keeps the wrong one, and that's much harder to catch than a raw WER miss because the sentence still reads fine. Do you know of anyone scoring disfluency-recovery as its own metric instead of folding it into WER? Feels like exactly the kind of thing that would need a downstream-task-aware eval rather than a transcript-only one.

by taiuo_ops

8/18/2026 at 2:16:35 AM

This looks really interesting.

I feel like there is a lot of room to build great voice-based agents that don't exist right now.

I have found that ChatGPT voice mode is unusable (e.g. hallucinates me saying things); Claude voice mode is usable, but very buggy around tool calling, and it often mishears things. And it only supports Opus, not Fable (though it looks like you don't support either of those). But I use it anyway.

Question, do any of your TTS options support increasing the speaking speed?

by dgreensp

8/18/2026 at 12:54:02 PM

> I feel like there is a lot of room to build great voice-based agents that don't exist right now

I agree. Even Codex’s new voice mode isn’t just “It does things I tell it”, all commercially available voice modes have to be dumbed down in the name of latency for some reason. I have even experienced the latency being uncomforably low before.

This is what I enjoy the most about my in-development personal OpenCode client: It’s just STT-LLM-TTS with the original model and all of its capabilities. Latency is obviously higher, but I don’t mind that much given the upside (would be interesting to optimize for completely blind usage, my current MO is mostly hybrid).

by solarkraft

8/18/2026 at 5:11:12 AM

thanks! yes, some of them do, cuz speed is a per-provider capability, not universal. And, you can see it in the gateway code (minimax, hume, xai tts adapters all handle a speed param). that unevenness is actually a routing constraint by itself: "voices that support rate control" narrows the candidate list the same way language or latency does. and agree on the voice modes, the gap between the demo and a dependable agent is exactly why we started this. what are you building with it?

by abdik

8/17/2026 at 5:19:03 PM

I made a completely free 100% on device translation app https://apps.apple.com/us/app/arda-translate/id6778970560 and hard to image a world where TTS and STT will not be done locally in the future

by cjjuice

8/17/2026 at 7:20:00 PM

cool app, and agreed that on-device keeps eating the single-user cases, we are seeing dictation and translation are exactly where local models shine. We benchmark the open models on the same boards as the hosted ones, but there are still a few: https://benchmarks.speko.ai/open.

production phone agents are a different shape today: the call terminates server-side, three models plus turn-taking under one latency budget, and per-language quality still swings a lot from our tests

by abdik

8/17/2026 at 5:55:54 PM

Which model best allows me to transcribe speech that uses a lot of domain-specific terms? For example, when I say "Claude Code", it often gets transcribed as "Cloud Code", and I have to go back and edit or do a second pass with a traditional LLM (which can introduce additional errors).

by maho

8/17/2026 at 7:23:11 PM

agree with omneity here. Whisper's initial-prompt trick is exactly that, and several hosted vendors have equivalents (custom vocabulary / keyword prompting). Domain vocabulary is where STT models separate the most in our runs. for example, on medical terms the field spreads from about 8% to 19% WER across models: https://benchmarks.speko.ai/blog/what-a-voice-agent-hears.

We often find that models that wins on clean speech are often not the one that wins on your terms, so test with your own vocabulary rather than a headline number.

by abdik

8/17/2026 at 6:15:16 PM

I’ve had a lot of success in the past with fine tuning STT using synthetic data.

I was doing it for Veterinary (ambient recording -> SOAP notes) which has tons of complex domain-specific language AND it is critically important to get right.

“CPR” transcribing as “see pee are” just doesn’t cut it in that industry.

by jeffrwells

8/17/2026 at 7:30:00 PM

Synthetic-data fine-tuning is the other credible answer to domain vocabulary. Curious whether you re-benchmark the fine-tune when new base models ship?

by abdik

8/17/2026 at 6:22:06 PM

Which open source STT models have you had success with for fine tuning?

by cootsnuck

8/17/2026 at 6:11:00 PM

Good old Whisper allows you to enter a prompt with domain specific terms and it will use them for transcription.

by omneity

8/17/2026 at 6:51:08 PM

Gemini 3 flash series is quite good, also elevenlabs scribe v2

by k9294

8/17/2026 at 5:39:55 PM

Ever since OpenAI launched their improved voice mode, I've been looking for a capable LLM with builtin voice-in and -out. Llama 4 was supposed to be it but turned out to be a dud. I haven't followed the topic closely lately, did I miss anything? Are there capable (!) open weights omni models that allow low latency voice chat? If so, what software do you use them with? Can you use a PWA on your phone? WebRTC? WebTransport?

by Tepix

8/17/2026 at 7:21:38 PM

Actually, we measured exactly this recently. The strongest open-weights speech-to-speech model we have run is NVIDIA's NemotronLabs VoiceChat 11B - no provider serves it, so we hosted it ourselves and ran the same scripted call every model on our board gets. Remarkably stable, 40+ sessions with zero errors - but by turn thirteen it was answering nine tries in ten without saying a word. Stable engine, but degrades on long calls.

Wider picture: we rebuilt the speech-to-speech test around a hard 20-turn concierge call, and today only the closed models finish it cleanly. Board: https://benchmarks.speko.ai/s2s

shared some write-up: https://benchmarks.speko.ai/blog/can-s2s-replace-the-cascade

And jakswa's Gemma datapoint matches what we measure, the fastest rows on our LLM board are the small models like Gemma 4 is performing incredibly well for voice agents

by abdik

8/17/2026 at 6:37:13 PM

FWIW, I have seen a couple phone apps using react native and sending directly to backends using a socket. Seems to avoid some overhead but I don’t have first hand experience with the client side.

by mlhpdx

8/17/2026 at 6:19:43 PM

Gemma 4 (both E4B + 12B) performed really well as ears+brains. I mostly comment because I too am always scouting for a nice local all-in-one model.

by jakswa

8/19/2026 at 3:55:11 AM

Came to check which open weights model is best right now for audio, found nothing, it seems you are just selling your services.

by siddbudd

8/17/2026 at 4:08:01 PM

What is the difference with Livekit Gateway? https://livekit.com/blog/introducing-livekit-inference

Or even something more managed like Vapi?

by MikhailTal

8/17/2026 at 4:25:46 PM

The main difference from gateway is we help with picking the right voice stack, which seems to be a big problem for users: we benchmark the models continuously and route based on those measurements for your language and constraints, and the boards are public at https://benchmarks.speko.ai/

Second difference is where it runs. Our gateway is open source and runs in your own container, including with self-hosted livekit/pipecat. You get a temporary token before the session starts, and then your orchestration connects directly to the provider.

Vapi is a managed platform: you use their infra to use the voice AI stack. In our case you can have your own infra and switch between models, so you are not locked into a vendor. A lot of teams we talk to build their own infra as they mature, and that is where the router comes handy.

by abdik

8/17/2026 at 4:24:13 PM

> Typical production voice agent is an ensemble of three models: STT, an LLM, and TTS.

To use a claudism, I would like to push back on this. The industry is very much moving towards one-model-does-all end to end trained similar to LLMs and VLMs. Mostly for latency reasons and partially because the results for the end to end trained models are just so much better than those using three pieces architectures.

I think most of the value prop is in automatic evals, not routing specifically. A better pitch for you would be "the LM Arena of voice models" rather than comparing yourself to openrouter because the value add is rather questionable. For TTS specifically, the current SOTA for production systems are all using prompt based voice gen i.e. instead of having 10 different Tacotron models trained on 10 different models, these days it's all a single large model and the "style" is a prompt in the system prompt. The input is usually something like

  <System prompt>
  Speak in a deep smooth voice similar to a documentary narrator
  </System prompt>
  <Text to Narrate>
  Speko is the ultimate evaluation platform for voice agents. We do automatic  evals.
  </Text to Narrate>
It's the same for voice cloning too, you just pass the reference speech as an input file for all generations. A lot of systems don't have any separate style vector extraction step or model-specific fine-tuning anymore.

So something like OpenRouter for voices offer questionable value given that stakeholders usually make this sort of decisions once at the start of the project. On the other hand if you can offer automatic evals and figure out which prompts give the most similar results across different voice providers, that would offer a lot more value. It would be nice to be able to switch from e.g. Grok voice agents to ChatGPT voice agents knowing that the output style won't change too much. There are many companies now with evals as a core business model: LM Arena, Artificial Analysis, Prompt foo (before they got acquired and pivoted to security only) so many take a look at them.

Source: we have been building TTS systems for over a decade too https://narrationbox.com

by narrationbox

8/17/2026 at 4:48:30 PM

Fair pushback. On end to end: we measure those too, same methodology: https://benchmarks.speko.ai/s2s. If the single models win, we route to them the same way, so we do not care which architecture (s2s or cascaded) wins. For now, what we see in production so far is that most teams still want to control each piece: swap the STT for medical vocabulary, keep the LLM, keep the voice.

On "promptfoo of voice models": that is closer to how it started. At my last company we ran these evals manually, we would even hire native-speaking raters, benchmark, switch if it wins. The evals are the value, agreed. The routing is what makes them actionable: teams told us swapping always looked like an R&D project, so scores alone did not change what ran in production.

On prompt-based voice gen and reference-audio cloning: agreed, that is what we see too. It makes continuous measurement more important: the same style prompt behaves differently per language and per content type, so we rank the voices themselves, tagged by use case: https://benchmarks.speko.ai/tts-voices

by abdik

8/17/2026 at 6:44:23 PM

> The industry is very much moving towards one-model-does-all end to end trained

I've worked with hundreds of enterprises on voice AI and voice agent solutions. In my experience, this isn't true. Or rather I should say, the people actually paying for voice agents (i.e enterprises) are not moving towards STS solutions in a meaningful way. The composability, observability, and reliability profile of STS systems is not amenable to enterprise criteria. Not to mention costs.

by cootsnuck

8/17/2026 at 7:31:02 PM

exactly, we see the same thing, around 95% cases are still cascaded, even tho STS has been improving a lot

by abdik

8/17/2026 at 10:32:11 PM

This is not true in my experience. By and large, everyone uses the cascading voice stack.

by jvwww

8/17/2026 at 4:58:41 PM

totally agree with this

by vdev123

8/17/2026 at 5:07:45 PM

Seems to be useless, the state of art for all categories is local on-device, voice model vendors are just rent seekers for those who know no better.

by bewareofscams

8/17/2026 at 5:09:23 PM

Just canceled my WisprFlow subscription a few days ago to switch to a open source, free, local alternative. In my case handy.computer with the Cohere model.

by sparkling

8/17/2026 at 5:10:57 PM

You mean handy.computer?

by alabhyajindal

8/17/2026 at 6:33:03 PM

Yup, my bad, fixed.

by sparkling

8/17/2026 at 8:15:14 PM

Open weight models are getting really good. Specko appears to be focussing on paid services which has its role to play but I have a benchmark in development for Open Weight models here if that is helpful: https://orcabot.com/benchmarks/open-weight-tts.

Still shoring up the details but you can try a sample of all the options and see how they compare in terms of model size, peak RSS, real time factor etc.

by robbomacrae

8/17/2026 at 5:39:42 PM

You're not the customer. This is for people building products that support thousands of users.

Linux on desktop is great for you, but this is a tool for people delivering solutions.

by echelon

8/17/2026 at 5:55:39 PM

> You're not the customer. This is for people building products that support thousands of users.

That's their dream. Your dream. The AI dream. Many would say it's AI psychosis.

by owebmaster

8/17/2026 at 5:42:09 PM

1) what if I told you I can leverage local models and serve thousands of users?

2) you know nothing about me

3) of course I am not! I do know better

by bewareofscams

8/17/2026 at 6:17:32 PM

What does WER/CER stand for? I see it as some kind of key metric under each model but not clear what it's measuring

by cnqso

8/17/2026 at 6:32:45 PM

Word Error Rate and Character Error Rate.

Common metrics to track how coarsely or finely accurate voice AI is

by purrcat259

8/17/2026 at 7:28:23 PM

What purrcat259 said, and I think that the page should explain it, we will add a tooltip.

for some languages CER is more relevant than WER. Thai and Mandarin have no word boundaries, so we score them by character, and Japanese gets a reading-based CER.

by abdik

8/17/2026 at 4:33:38 PM

Looks awesome, can't wait to try it for some Filipino workflows when it's available!

by dayvough

8/17/2026 at 4:58:18 PM

thanks! actually, we have the filipino already, can you check out and share your feedback?

by abdik

8/17/2026 at 9:01:35 PM

Is this only for realtime tts use cases?

Wondering if you also support some non realtime models.

by aleksiy123

8/17/2026 at 9:35:45 PM

yes, it does support, when you are creating an api key, you can point out narration or transcription use case, then you will be able to see. let me know how it goes or what use cases are thinking of for non-realtime models?

by abdik

8/18/2026 at 4:16:42 AM

why isn't openrouter gonna be the openrouter for voice ?

by alexcnwy

8/18/2026 at 4:37:13 AM

People who built or building voice agents immediately get this.

openrouter is the openrouter for audio models.

the conflation is "audio models" vs "voice ai", and the mental model that untangles it: think batch requests. text in, audio out (streamed, even). audio in, transcript out.

three questions inside the word "router":

1 - what gets picked: a model/voice, a provider of the same model, or the whole stack (stt + llm + tts) per call

2 - where it lives: an external http gateway, the agent platform (vapi/retell/livekit configs), or inside the live session

3 - when: session start, or mid-call

a voice agent is not a batch request. it's a live duplex session: turn-taking, barge-in, telephony legs, session state. the latency physics diverge too: a middleman hop in the media path is paid once by a batch request and on every conversational turn of a live call, so the media path wants a direct connection to the provider. a gateway that terminates at http can route the requests inside a call; it can't route the call.

by abdik

8/18/2026 at 4:39:57 AM

OpenRouter isn't open source.

This isn't Stripe payments. The market will have lots of competitors.

by echelon

8/17/2026 at 4:12:02 PM

the concept is interesting I must say - good luck !

by dhruv3006

8/17/2026 at 4:14:53 PM

Thank you!

by abdik

8/17/2026 at 6:07:02 PM

any progress with on-device models??

by IgorBlink

8/17/2026 at 6:19:48 PM

there is a good progress on on-devise models, but not ready for production yet to fit in devices. But as soon as there is are some good results, we are going to benchmark them and put in https://benchmarks.speko.ai/

by abdik

8/17/2026 at 7:24:47 PM

cool!

by IgorBlink

8/18/2026 at 7:04:17 AM

[dead]

by AITokenHub

8/17/2026 at 4:04:38 PM

The link, since it seems to be missing?

https://speko.ai/

by greyb

8/17/2026 at 4:13:40 PM

Yes, i added it. that's the right link.

by abdik

8/17/2026 at 4:55:23 PM

Awesome. Thanks for sharing!

by greyb

8/18/2026 at 2:49:07 AM

[flagged]

by Taikhoom10

8/18/2026 at 5:11:58 AM

i don't have a thesis on voice as a form factor for search. the demand we serve already exists: businesses answer phones. receptionists, outbound campaigns, clinic front desks and these calls happen at scale today, and the teams running them are the ones picking speech models. whether voice wins new interfaces is a separate bet

by abdik

8/18/2026 at 1:05:22 PM

So how exactly are you counter-positioned to Eleven Labs? They are serving a similar market correct?

by Taikhoom10

8/18/2026 at 7:00:31 PM

not really counter-positioned, they are one of the providers we route to. elevenlabs sells models - voices, stt, and their agent platform on top of them. we don't sell any models: we measure all of them (elevenlabs included, their scribe is near the top of our stt board) and route each request to whatever wins for your language and constraints. so for them the best outcome is that you use their models; for us the best outcome is that you use the right one. sometimes that is the same thing.

by abdik

8/18/2026 at 10:13:43 PM

ah so literally OpenRouter for voice eh? Well then I would ask, what happens when OpenRouter just ads voice, seems simple right?

by Taikhoom10