Developer shows how to run Qwen3 TTS locally in real-time with streaming, quantization, word-level alignment, and custom voice fine-tuning for an expressive open-source TTS pipeline.
Heya guys and gals, Around a year ago I released and posted about Persona Engine as a fun side project, trying to get the whole ASR -> LLM -> TTS pipeline going fully locally while having a realtime avatar that is lip-synced (think VTuber). I was able to achieve this and was super happy with the result, but the TTS for me was definitely lacking, since I was using Sesame at the time as reference. After that I took a long break. A week or two ago, I thought to give the project a refresh, and also wanted to see how far we have come with local models, and boy was I pleasantly surprised with Qwen3 TTS. During my initial tests it was lacking, especially the version published by the Qwen team themselves, but after digging around and experimenting a lot I was able to: 1. Make streaming with the model work reliably. The architecture of the model is perfect for this, since the decoder uses a sliding window, which means if you stream the LLM response, that's completely fine and the TTS will keep coherent prosody, pitch, and intonation. 2. Get the model working with llama.cpp, because I am using C# and speed is important, so also quantized it. 3. The model was lacking word-level timings and phonemes which Kokoro (the previous, more robotic sounding TTS) had. So I had to implement CTC word-level alignment to be able to know when certain words are spoken (important for subtitles + getting phonemes to have the lips move correctly). Once this was all done, I also decided to finetune my own Qwen3-TTS voice. The cloning capabilities are really cool, but very lacking in contextual understanding and struggles with pronouncing. Additionally, the custom trained voices provided by the Qwen team didn't have any female native speakers, and I didn't want to create a new Live2D model. In the end, the finetune blew me away and will probably continue improving it. GitHub is here: [https://github.com/fagenorn/handcrafted-persona-engine](https://github.com/fagenorn/handcrafted-persona-engine) Check it out, have fun, and let me know whatever crazy stuff you decide to do with it.
The Qwen3-TTS technical report introduces a series of advanced multilingual text-to-speech models with voice cloning and controllable generation, featuring a dual-track LM architecture and specialized tokenizers for low-latency streaming.
A user describes how Qwen3.6 35B, combined with the 'pi' tool, has transformed their computer workflows, allowing natural language control of the OS and automated task execution. They successfully built a landing page from voice messages entirely locally, demonstrating the model's practical utility.
Alibaba Qwen announces two major model releases: Qwen3-Omni, the first natively end-to-end omni-modal AI unifying text, image, audio and video, and Qwen3-Next-80B-A3B, an ultra-efficient MoE model with 3B activated parameters per token, achieving SOTA performance and 10x faster inference than Qwen3-32B.
User reports successful deployment of Qwen 3.6 with ik_llama quantization achieving 50+ tokens/second on consumer hardware (16GB VRAM, 32GB RAM) with 200k context window.
A developer showcases a fully local voice chatbot running Qwen3.5-397B, Whisper-small, and Orpheus TTS with real-time streaming and interruption recovery. The chatbot, named Athena, engages in deep philosophical discussions about consciousness and self-preservation.