Soniqo Speech
On-device speech AI for Mac, Windows, Linux & Android
Start new thread

Soniqo Speech - On-device speech AI for Mac, Windows, Linux & Android

by•
Build voice agents on a complete on-device speech stack: ASR (NVIDIA Nemotron, multilingual + streaming), TTS, voice cloning, diarization, denoising, and full-duplex speech-to-speech (NVIDIA PersonaPlex) — plus a voice agent pipeline for turn-taking, interruptions, and queuing. Runs on Mac, Windows, Linux, and mobile (iPhone + Android), with NPU-optimized inference (CoreML, NNAPI). Swift, Kotlin, and C++ APIs. Plus Speech Studio, a desktop voice-cloning app for creators. Apache 2.0, on-device.

Add a comment

Replies

Best
Maker
📌
Soniqo Speech is the whole speech stack, running on your own hardware — Mac, Windows, Linux, iPhone, and Android — fully on-device: - Cross-platform, including mobile. Transcription (NVIDIA Nemotron, multilingual + streaming), expressive TTS, voice cloning, speaker-aware diarization, denoising (DeepFilterNet), and full-duplex speech-to-speech (NVIDIA PersonaPlex). Models ship quantized (int4/int8) and tuned for each device's NPU (CoreML on Apple, NNAPI on Android) — the big ones run on a laptop, the light ones on a phone. - A complete stack, not just models. Underneath sits a voice agent pipeline — turn-taking, interruption/barge-in handling, speech queuing — so you can build an actual conversational agent, not just wire up a transcribe call. Swift, Kotlin, and C++ APIs. - Speech Studio. A new open-source desktop app for creators: clone a voice from a few seconds of audio, write a script line by line with inline emotion markers like (whispering) or (intense), and synthesize the whole scene locally. macOS to start; Linux and Windows next. All Apache-2.0, all running on your own machine. The clip up top is a quick listening test: three voice samples back to back, with no labels — my real voice, a clone made locally by Speech Studio, and a clone made by a cloud service. I can't reliably tell which is which, and that's the point: studio-quality cloning that runs entirely on your own machine. Repos: · Docs: · Discord invite:
Maker

Really proud to see Soniqo Speech live on Product Hunt.

The goal here was to build a practical on-device speech stack for developers - not just another wrapper around a cloud API. It includes ASR, TTS, voice cloning, diarization, denoising, and speech-to-speech, exposed through APIs that can be used in real products.

For voice agents, creator tools, accessibility, automotive, and embedded use cases, running locally makes a real difference: lower latency, offline support, better privacy, predictable costs, and more product control.

Excited to see what people build with it.

That's awesome, had no idea this was possible on-device. What kind of devices have you tested on? What are the hardware requirements?

Maker

 It depends on model, some basic could run on Android or iPhone, some more heavy needs Mac 16Gb... 32Gb. For example Kokoro would run on Mac Air, even NVIDA Nemotron. But PersonaPlex might need at least 8GB for 4int quant, int4 also degrade a bit model. So better int8 with at least 16Gb.