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?
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.
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.
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.
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.
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.
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.
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/
> 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.
> 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.
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
https://benchmarks.speko.ai/turntaking
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.
Common metrics to track how coarsely or finely accurate voice AI is
Linux on desktop is great for you, but this is a tool for people delivering solutions.
2) you know nothing about me
3) of course I am not! I do know better
That's their dream. Your dream. The AI dream. Many would say it's AI psychosis.
Or even something more managed like Vapi?
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.
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
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
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.
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
https://speko.ai/