NeuTTS-2E
NeuTTS-2E is a super-fast, highly realistic, on-device emotional TTS speech language model. It is an early alpha release, English-only model, supporting six emotions plus neutral (`angry`, `disgusted`, `fearful`, `happy`, `sad`, `surprised` and `neutral`) across four fixed speakers (`emily`, `paul`, `sophie`, `steven`). With a compact backbone and an efficient LM + codec design, NeuTTS-2E delivers strong naturalness and expressive control at a fraction of the compute, making it ideal for embedded agents, games, robotics, toys, and offline assistants. The model processes text, speaker, and emotion inputs to generate speech in real-time without GPUs, enabling private, low-latency voice output on consumer hardware. NeuTTS-2E is open source with available pip install (including ONNX runtime option) and a Hugging Face Space for live demos. It addresses the need for controllable, emotionally expressive speech in resource-constrained environments where cloud APIs introduce latency or privacy concerns.
Reader rating
No ratings yet
You might also like
Related tools
Earshot is an open-source toolkit that gives coding agents practical access to audio understanding and sound creation. It ships as a command-line tool, Python API, and Model Context Protocol server, allowing compatible agents to inspect recordings, search for relevant moments, compare audio against text, and generate or fit procedural sound recipes. Developers building audio, video, or creative applications can use it to make sound part of an agent workflow instead of relying only on transcripts or external hosted services. Earshot is designed to run locally on CPU, which keeps media private and avoids per-request cloud inference, though its documented models require a substantial initial download. Its recent launch visibility reflects growing demand for multimodal agents that can work with sound as well as code and text.
Desert Ant Labs provides a family of small, specialized AI models and native SDKs designed to run on-device instead of through metered cloud inference. Its product line includes Voz for speech recognition, Clear for audio enhancement, Redact for privacy filtering, Clips for selecting video highlights, and other focused models that can be embedded into apps with Swift, Kotlin, or JavaScript. Voz is a concrete example: it transcribes 25 languages with word-level timestamps, runs on Apple platforms, and can process long audio locally with no login or per-token bill. The service is aimed at mobile and desktop developers building private, responsive AI features into their own products. Its official site offers free usage up to 100,000 monthly active devices per platform, while documentation and Hugging Face model pages provide installable evidence. This is an SDK/model platform, not merely a research announcement.
Demon (Diffusion Engine for Musical Orchestrated Noise) is an open-source real-time music generation system that runs locally on consumer GPUs at 25Hz. It is built for musicians, sound designers, music producers, and AI researchers who want to generate, iterate, and perform with AI music in real time without relying on cloud APIs. The system uses diffusion-based synthesis to produce musical audio streams with low latency, enabling live experimentation and performance workflows. Demon launched on Hacker News with 15 points and the project page at daydreamlive.github.io/DEMON describes a fully local, GPU-accelerated approach to music generation. What makes it notable is the combination of real-time performance with diffusion models — a technical achievement that opens up live music creation use cases that were previously impossible with slower batch-generation approaches.