Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

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

speko.ai

Post reply on HN
Print view View original

41–50 of 73 posts

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

#41

Post by abdik »

cjjuice wrote: post #21

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

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

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

#42

Post by abdik »

Tepix wrote: post #23

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? WebTrans…

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

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

#43

Post by abdik »

maho wrote: post #26

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).

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.

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

#44

Post by IgorBlink »

abdik wrote: post #32
IgorBlink wrote: post #27

any progress with on-device models??

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/

cool!

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

#45

Post by abdik »

cnqso wrote: post #30

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

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.

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

#46

Post by abdik »

jeffrwells wrote: post #29
maho wrote: post #26

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).

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.

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?

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

#47

Post by abdik »

cootsnuck wrote: post #37
narrationbox wrote: post #7

> 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 m…

> 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…

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

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

#50

Post by robbomacrae »

bewareofscams wrote: post #18

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.

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.

Post reply on HN