AWS ships Qwen3-TTS-12Hz-1.7B-Base voice cloning on SageMaker JumpStart

The Qwen3-TTS family, developed by the Qwen team at Alibaba Cloud, uses the Qwen3-TTS-Tokenizer-12Hz speech tokenizer and supports streaming generation for low-latency, interactive scenarios. The Base variant used here performs voice cloning from a few seconds of user audio and can also serve as a base for fine-tuning. It differs from the CustomVoice variant, which generates speech from a fixed set of predefined speakers rather than a user-supplied reference.

The deployment runs the model as two stages on the same GPU: a talker stage that generates speech tokens from text and reference voice, and a code2wav stage that renders them into a waveform. At container startup, each stage logs memory use to the endpoint's CloudWatch Logs log group. The talker model weights are 3.66 GiB and the code2wav model weights are 0.45 GiB, with most of the remaining memory becoming KV cache — the working memory that holds tokens for in-flight requests.

The key test for this deployment is whether teams can keep the two stages' combined GPU memory reservation within the 24 GB capacity while still handling enough concurrent requests. The article notes that one instance can accommodate multiple simultaneous requests because TTS requests are typically short, but for higher traffic, more instances or a larger GPU would be needed. The walkthrough shows how to manage that balance with configuration and monitoring, which is where the practical value of this news lies for teams building voice-enabled applications.