Toby Kim (@_doyeob_) header
Toby Kim (@_doyeob_) profile picture

Toby Kim

@_doyeob_

Founder of @trynarilabs: multimodal inference at the speed of light.

 San Francisco
narilabs.com
Joined September 2023
  • Tweets92
  • Following328
  • Followers6.2K
19Media
Photo by Toby Kim (@_doyeob_)Photo by Toby Kim (@_doyeob_)Photo by Toby Kim (@_doyeob_)Photo by Toby Kim (@_doyeob_)Photo by Toby Kim (@_doyeob_)Photo by Toby Kim (@_doyeob_)Photo by Toby Kim (@_doyeob_)Photo by Toby Kim (@_doyeob_)Photo by Toby Kim (@_doyeob_)
Two undergrads. One still in the military. Zero funding. One ridiculous goal: build a TTS model that rivals NotebookLM Podcast, ElevenLabs Studio, and Sesame CSM. Somehow… we pulled it off. Here’s how 👇
217
563
151
5,307
727,130
Toby Kim retweeted
Yurii Rebryk (@rebryk) profile picture
Yurii Rebryk
@rebrykOct 27
Accuracy doesn't have to cost speed. The two most accurate systems return final text 0.08 s (MAI-Transcribe-2) and 0.14 s (Scribe v2 Realtime) after the speaker stops. Only four systems are on the Pareto front, where nothing is both faster and more accurate.
Image from Yurii Rebryk (@rebryk)
1
1
39
1,163
Great to see our Qwen3-ASR 1.7B Fast endpoint at #3 of 21 here 🎉 > 6.84% WER across all accents > 70 ms median latency > on the speed/accuracy Pareto front Only ElevenLabs and Microsoft are ahead of an open-weights 1.7B model. That's Qwen running on our inference engine 🔥
Yurii Rebryk (@rebryk) profile pictureYurii Rebryk
@rebrykOct 27
Speech-to-text is solved. Unless you have an accent. Over 1 million people from 180 countries have spent 580,000 hours speaking with Fluently AI tutors while learning a new language. Nearly all of them have an accent, and transcribing their speech is still hard. So we benchmarked streaming speech-to-text systems on spontaneous, accented English from our learners. Leaderboard and methodology in the thread. English is the easy one. Our learners study 14 other languages too. Working on tough speech problems? DMs are open.
4
4
1
53
4,017
Nemotron VoiceChat from @nvidia listens and responds in audio while also calling external tools to retrieve information or perform tasks. We build an inference engine that handles 64 sessions on a single H100 GPU while handling tool-calls effectively. blog: narilabs.com/blog/making-too…
Image from Toby Kim (@_doyeob_)
3
5
2
100
5,317
vLLM-omni does 1 session per GPU. NVIDIA's own NIM and C++ inference engines are each capable of 4 and 6 concurrent sessions on 1x H100. At 64 sessions, the cost to run S2S sits at around 6.7 cents per hour, per user.
1
5
407
Amazing feature from Speko. Great to see our Qwen3-ASR endpoints sitting at #1 and #2!
Speko (YC S26) (@speko_ai) profile pictureSpeko (YC S26)
@speko_aiSep 28
Stop guessing which speech model is fastest or cheapest. Pick up to 3, set your language, and see latency, price per minute, and capabilities next to each other. Speech to text, text to speech, and language models. Try it: platform.speko.ai
2
2
16
2,178
The video doesn’t seem to show, but here’s what shows up on the platform.
Image from Toby Kim (@_doyeob_)
2
158
Our TTS and STT endpoints are now live at @speko_ai: Openrouter for voice AI We're sitting at #1 in latency in both STT and TTS on Speko benchmarks. And our STT endpoint is top 2 in terms of accuracy beating out Deepgram, Meta, and Gemini. Try it out: speko.ai
Image from Toby Kim (@_doyeob_)
4
2
1
48
4,126
heavily requested feature: input streaming for Qwen3-TTS is now GA! stream your LLM outputs into our TTS endpoint so you can output audio before the LLM is finished ✅
Image from Toby Kim (@_doyeob_)
2
2
1
37
2,867
blog: narilabs.com/blog/input-stre…
Input Streaming for Nari Qwen3-TTS

Input Streaming for Nari Qwen3-TTS

Send text as your LLM generates it and receive speech from the same WebSocket before the full response is ready.

narilabs.com
4
331
Just added word timestamps support to our Qwen3-ASR streaming endpoint - using the 0.6B forced aligner model. Still the same latency and cost! Blog: narilabs.com/blog/word-level…
Image from Toby Kim (@_doyeob_)
1
3
2
42
2,851
Our STT and TTS endpoints are entering GA today 🥳 Try now and receive welcome credits for testing! > TTS Fast: $10 / 1M (top TTFA, 50 ms) > TTS Standard: $5 / 1M > STT Fast: $0.12 / hr (top TTFS, 40 ms) > STT Standard: $0.06 / hr
Image from Toby Kim (@_doyeob_)
8
3
2
63
3,200
try now at narilabs.com
Nari Labs — Multimodal Inference at the Speed of Light

Nari Labs — Multimodal Inference at the Speed of Light

Ship multimodal models with lower latency and higher efficiency. Explore optimized model APIs, dedicated enterprise inference, and fine-tuning with Nari Labs.

narilabs.com
1
322
Duplex speech is the frontier of Speech AI and PersonaPlex 7B from @nvidia is the most adopted OSS duplex model. We achieve 80 concurrent sessions on a single H100 GPU, costing below 10 cents an hour per user to serve. GPT-Live costs 10 cents per 2 minutes. Read more: narilabs.com/blog/personaple…
Image from Toby Kim (@_doyeob_)
12
10
1
179
8,625
There's still a long way to go before OSS models reach the intelligence and naturalness of models such as GPT-Live. But we are bullish about the future of open duplex models and will continue to build infra to serve them to the masses, cheaply.
2
7
726
Fun fact: PersonaPlex's original experimental setup used our Dia TTS model to synthesize its two-speaker customer-service conversations.
2
5
491
Diarization is a key problem in speech. And Pyannote is the king of diarization. Used it for all of Dia's data pipelines. Through various optimizations, we got Pyannote's flagship OSS model to process 100+ hours of audio in under 100 seconds. on a single A100 GPU. Read more: narilabs.com/blog/making-pya…
Image from Toby Kim (@_doyeob_)
3
3
1
94
4,321
Nari Labs' Qwen3-TTS and Qwen3-ASR endpoints lead @covaldev voice AI benchmarks 🎉 Text-to-Speech > #1 Lowest Latency (time-to-final-segment): 44 ms > #2 Accuracy: 3.6% WER, just 0.1% shy of #1 Speech-to-Text > #1 Accuracy: 3.8% WER > #2 Lowest Latency (time-to-first-audio): 63 ms Read more: narilabs.com/blog/nari-labs-…
Image from Toby Kim (@_doyeob_)
6
6
1
69
5,516
Did I mention prices? > STT: #2 cheapest > TTS: tied for #1 cheapest Try now at narilabs.com for FREE. Benchmark values are based on Sep 14 15:00 UTC, only comparing publicly available endpoints.
Nari Labs — Multimodal Inference at the Speed of Light

Nari Labs — Multimodal Inference at the Speed of Light

Ship multimodal models with lower latency and higher efficiency. Explore optimized model APIs, dedicated enterprise inference, and fine-tuning with Nari Labs.

narilabs.com
2
383
Today we’re launching: > the world's fastest Speech-to-Text endpoint at 40 ms time-to-final-segment (TTFS) > the world's cheapest streaming Speech-to-Text endpoint at $0.06 / hr Powered by Qwen3-ASR 1.7B on Nari Labs inference engine. 🔠 Top 3 Accuracy on Coval Benchmarks 🚀 3x faster than ElevenLabs Scribe V2 💸 9x cheaper than Gemini Transcribe 3.5 ✅ available FREE for a limited time We believe open-source will win - not just in LLMs, but also in multimodal AI. Nari Labs is here to accelerate that future, starting with speech. Try now at narilabs.com
27
35
1
414
23,991
more replies
We are a small team of ex-YC, ex-KRAFTON research engineers. We built Dia, the first OSS model for natural dialogue w/ 2M+ downloads and ~20k Github Stars. Our team has published on ICLR / NeurIPS and earned 3 IOI and ICPC gold medals. @trynarilabs is backed by @ycombinator We’re just getting started. More to come in Speech, Audio, Video, and World Models. Stay tuned. narilabs.com
Nari Labs — Multimodal Inference at the Speed of Light

Nari Labs — Multimodal Inference at the Speed of Light

Ship multimodal models with lower latency and higher efficiency. Explore optimized model APIs, dedicated enterprise inference, and fine-tuning with Nari Labs.

narilabs.com
2
1
6
3,275
Happy to see us on top of Coval Benchmarks!
Image from Toby Kim (@_doyeob_)
Image from Toby Kim (@_doyeob_)
511
Load more