


PhD | AI & voice | startups linkedin.com/in/josh-r-meyer…


Are You Smarter than a 5th Grader, AI edition episode 1: Haru-4B
i tried out moshi locally on my macbook. moshi is the OG speech-to-speech model... and another fun conversation :)

Introducing Trelis/Duplex-0 huggingface.com/Trelis/Duple… An Apache 2-licensed speech-to-speech prototype model and training scripts. Duplex-0 is a ~200M parameter duplex model, trained on ~750 hrs of audio. It is a research checkpoint that is useful for ablating data mixes and model architectures... its speech is about half as intelligible as real speech in our measurements. Duplex-0 comes in three architectural flavours: 1. two-track audio 2. two-track audio + 1-track model text 3. 2-track audio + 1-track model text + 1 track caller text (ASR, 320 ms delay) These three training options were compared on the same data, seed and settings. All three produce speech of similar quality: 60 to 68 confirmed words per minute after the conversational stage, against 132.8 for real speech. Here are some further notes on its design and training.

You ask a question, the line goes quiet, and then comes the “hello?... oh sorry, go ahead.” @veris_ai measured that pause across 300 simulated billing calls, and Dialog-RSN-1 replied the fastest of every voice agent on its board. @matthen2 breaks it down: bit.ly/3WT55Ai

🫶
thanks for testing it out! we should have an improved model out soon too :D

You ask a question, the line goes quiet, and then comes the “hello?... oh sorry, go ahead.” Veris AI measured that pause across 300 simulated billing calls, and Dialog-RSN-1 replied the fastest of every voice agent on its board. @matthen2 breaks it down: bit.ly/3WT55Ai

🔗 @veris_ai write-up: veris.ai/blog/vamos-pro-poly… 📊 Full leaderboard: veris.ai/leaderboard/vamos-p… 📝 PolyAI's post: poly.ai/blog/veris-dialog-rs… 🎙️ Dialog-RSN-1: poly.ai/dialog-rsn-1

PolyAI says Dialog-RSN-1.1 will ship very soon 👀 should have closer instruction following, stronger Spanish, French, German, Dutch, Portuguese, Arabic, Japanese and Mandarin... and speaking while a tool call runs 🛠️ 🧵/

so... how is @polyaivoice's Dialog-RSN-1 so fast? the model hears raw audio and responds directly in text, then a separate fast TTS speaks it no intermediate transcript needed 👀 🧵/

what is VAmoS Bench by @veris_ai ? a voice agent has to answer the billing line of an electric utility ✅ 300 phone calls ✅ calm + angry callers ✅ quiet room / street / café / TV in the background same calls, same tools, same instructions for every agent on the board 🧵/

🔥 NEW fastest voice agent on VAmoS Bench 🔥 @polyaivoice's Dialog-RSN-1 replies 1.24 s after the caller stops talking (median). Most other agents take over 2 s ⏳ details in thread 👇 🧵/

Small CPU TTS bench updates - Now with UTMOS - PiVo-1.4M: 5 ms to first audio, 9x real time - Inflect-Micro-v3: 11x real time, 0.8% WER - Supertonic 3: UTMOS 4.47 Kokoro-82M missed real time. erogol.com/small_tts_bench/

I really like their blog! Some thoughts: 9M hours at 12.5Hz taking ~180k steps => assuming bf16, they trained on approx 40-50 TB VRAM (3-4 GB200 racks). This is 13 TPP in total, or 135 TPAP. Plateau at ~3.6M hours and 31B a3B, napkin math gives ~16 active experts per layer

.md would've been more practical than print
Ben Arledge @BenArledge Nov 17

