Tym Rabchuk

AI Developer
Offsec Lead

I build and test various STT and TTS models and design voice agents

[Tym Rabchuk @5uck1ess]

@pablogimenezz amazing. if you get a chance, listen to a bunch of them. id love to get a professional opinion since it's not my native language, especially on accents

[Tym Rabchuk @5uck1ess]

just shipped: tts-bench now benches 28 paid tts apis head to head 🔥

same prompts as the 89 open models. speed, samples, scores, price, all on one page

fastest: @AssemblyAI hits first speech in 30ms. @GradiumAI 53ms, @PalabraAI 77ms
best sounding: @VUILabsAI luna. it also cloned my voice the best out of 21. but it takes 1.2s to start talking

go listen and pick your winner 👇
https://t.co/Qf3k68a7JL

[Tym Rabchuk @5uck1ess]

@pablogimenezz already in there: prompt 5 is french and prompt 6 is spanish, every api has samples for both. you'd hear it better than me, curious which ones hold up in spanish

https://t.co/aOQ0SPycfx
https://t.co/jdumW2Tznd

[Tym Rabchuk @5uck1ess]

@AssemblyAI @GradiumAI @PalabraAI @VUILabsAI https://t.co/jdumW2Tznd

[Tym Rabchuk @5uck1ess]

@AssemblyAI @GradiumAI @PalabraAI @VUILabsAI everyone else on the board: @ElevenLabs @cartesia @inworld @SpeechifyAI @FishAudio @MistralAI @rimelabs @soniox_ai @StepFun_ai @smallest_AI @DeepgramAI @MicrosoftAI @OpenAI @GeminiApp @HeyGen @MiniMax_AI @usebland @resembleai. samples for all of you are up, have a listen

[Tym Rabchuk @5uck1ess]

@theo whoever is not using it, is kinda missing out. perfomance is good and phone remote control has never been better. finally a single tool to manage them all

[Tym Rabchuk @5uck1ess]

Claude Duet is getting better

@gabemulley that's a great add. finally some integration with Duet and Claude Code.

@LLMJunky exactly what we complained about.

[Tym Rabchuk @5uck1ess]

Is @Muse really the only always-on agent that has Android messages access?

[Tym Rabchuk @5uck1ess]

@theo Just try out Muse. Theo is right. conversation does feel good. it's hard to verbalize. I just feels right.

[Tym Rabchuk @5uck1ess]

@LLMJunky I thought this wasnt a big deal but it is. the integration is missing that Dots has. One Duet channel has no idea about claude code doing or anything. Plus github integration is so weird. it's all via threads not via single central integrator.

[Tym Rabchuk @5uck1ess]

@soniox_ai this is already one of the better STT, this is incredible. Great work!

[Tym Rabchuk @5uck1ess]

@jack would love to put dessa on tts-bench next to the 28 hosted apis once devs can try it. is the 16 ms from real end of speech or from when it predicts you're done?

[Tym Rabchuk @5uck1ess]

@yasser_elsaid_ which voice-to-voice model is doing the talking? your own or something like gemini live / gpt-realtime underneath? very curious

[Tym Rabchuk @5uck1ess]

@NicklausSalzman @rimelabs amazing ty. rime latency is very good.

[Tym Rabchuk @5uck1ess]

@meetgranola I hope this STT engine is good. Wispr flow has been very nice but if you can beat that, I'll be very glad to switch. Granola was my first product for work meeting notes.

[Tym Rabchuk @5uck1ess]

@Muse @alexandr_wang I understand what @theo now means about Muse conversation feeling just right. it's hard to put it in words but it mirrors your conversation style like no other agent

[Tym Rabchuk @5uck1ess]

Finally got access to live voice @Muse. Very impressive @alexandr_wang. I think this is the new winner for conversation quality compared to Dots, Grok Bot, and Duet (background voice FINALLY enabled)

Duet is slowly climbing up in my rankings and is more capable.

Muse speech is top notch though. speaks in tldr which is amazing! I can't stress enough that Muse's STT is on par with Wispr flow from all my tests. this is very impressive.

[Tym Rabchuk @5uck1ess]

@faye04081 @SamuelZengML @Edge0_AI thank you! haven't measured concurrency, the bench is meant to be an easy python setup to get models running and tested. for proper concurrency i'd use something like audiocpp (i contribute there too)