Samuel Zeng

Founder, | Building the voice layer for AI: cloud speech foundation models + on-device intelligence | MIT TR35

[Samuel Zeng @SamuelZengML]

The future of voice is hybrid.

I'm building https://t.co/DL4dsiRCFF around a stack: cloud speech foundation models plus on-device intelligence.

Cloud brings quality and scale. Devices add privacy, low latency, and reliability.

The goal: make voice native to every AI system.

[Samuel Zeng @SamuelZengML]

demo page: https://t.co/8VFAGc3JVN

[Samuel Zeng @SamuelZengML]

Audio8 TTS Preview 0.6B ONNX INT4 is now available. A compact model with serious capability:

- Runs entirely on CPU with ONNX Runtime
- Uses only around 1 GB of memory during inference
- Multilingual TTS across 11 languages
- Zero-shot voice cloning
- No PyTorch or Transformers required
- Includes CLI, streaming, HTTP/OpenAI-compatible APIs, and voice registration Small footprint. Full voice-cloning capability.

Small footprint. Full voice-cloning capability.

Model: https://t.co/lElyitoZlx
Code: https://t.co/1sKAvqDNxY

[Samuel Zeng @SamuelZengML]

Audio8 ASR is part of a broader thesis: high-quality speech intelligence should be open, multilingual, and deployable across cloud and edge.

Paper:
https://t.co/edZYdLBIHD

Try the models, star the repo, and tell us what you build.

[Samuel Zeng @SamuelZengML]

Open-source ASR just took the top spot.

Audio8 ARK-ASR-3B now ranks #1 on the @huggingface Open ASR Leaderboard with a 4.76 Mean WER.

Our 0.6B model also ranks in the Top 5.

Two model sizes. One open speech stack. Thread ↓

[Samuel Zeng @SamuelZengML]

Training, inference, evaluation, and vLLM serving code are open.

The repo includes Transformers inference, batch evaluation, a native /asr endpoint, and an OpenAI-compatible /v1/audio/transcriptions endpoint.

GitHub:
https://t.co/LOdRIwZA7c

[Samuel Zeng @SamuelZengML]

Open source means more than a checkpoint:

• weights + bundled neural codec
• zero-shot/no-reference inference
• JSONL batch generation
• codec-index preparation
• single-/multi-GPU SFT

Build with it. Break it. Tell us where it fails.

https://t.co/2NbwW5AnTe

4/4

[Samuel Zeng @SamuelZengML]

A 0.6B speech model should not sound this capable.

Today, we’re open-sourcing Audio8-TTS Preview:

11 languages.
Zero-shot voice cloning.
A bundled 44.1 kHz codec.
Apache 2.0.

Small model. Serious speech.

Thread ↓

[Samuel Zeng @SamuelZengML]

Benchmarks aren’t enough for speech. Hear it.

The listening set covers multilingual speech, pronunciation challenges, tongue twisters, classical Chinese, hard cases, and cross-lingual voice cloning.

Listen: https://t.co/Xljm91vbaZ

3/4

[Samuel Zeng @SamuelZengML]

For reproducibility, the model card reports a 7.03 WER mean across the seven current Open ASR Leaderboard splits. The dataset revision, inference setup, and raw manifests are documented here: https://t.co/nuwzrjqTz5

[Samuel Zeng @SamuelZengML]

We just open-sourced Audio8-ASR-0.1B, one of the smallest production-ready ASR models with cloud-level accuracy.
Only 100M parameters, under 200MB peak memory, and built to run on real hardware. No cloud required.
https://t.co/nuwzrjqTz5

[Samuel Zeng @SamuelZengML]

From meeting notes and live captions to voice assistants and smart hardware, we want high-quality ASR to run directly on every device, without being limited by cloud access or compute.
Small model. SOTA-level performance. Fully on-device.
GitHub: https://t.co/eyt0zrmWD8
Paper: https://t.co/B6sN872lpV

[Samuel Zeng @SamuelZengML]

Built for on-device deployment:
✅ ONNX Runtime + Apple Neural Engine
✅ Local transcription with low latency and high privacy
✅ 7 languages: Mandarin, English, French, German, Japanese, Korean, and Cantonese
✅ Hotword boosting without fine-tuning

[Musolsol.𝟎𝐱𝐔 @MMMusol]

网易云花了十年都没做出来的歌词体验，
被一个独立开发者用开源项目实现了

Folia —— 音乐可视化+动画歌词，丝滑程度头皮发麻

不夸张地说，这是我见过审美最在线的音乐类开源项目，没有之一，UI 细节拉满

强烈推荐给所有对音乐播放体验和审美有追求的朋友

GitHub：https://t.co/NN163FLr4y

[Samuel Zeng @SamuelZengML]

🚀 Just open-sourced OpenResearch!

AutoResearch + AlphaEvolve-style local search for code, prompts & training loops.

Key features:
• Mission-driven autonomous iteration
• Persistent memory + fully replayable artifacts
• Adapters for prompt eval, benchmark tuning & real GPU training
• Beautiful HTML reports & proposal journals
• 100% local, uv one-click, works with any OpenAI-compatible LLM

→ https://t.co/VdEXTDlk8f

Star ⭐, try it, contribute! Let's automate research together 🔥

#OpenSource #AI #LLM #AutonomousAgents #MachineLearning