AWS published a guide for deploying the Qwen3-TTS voice cloning model on SageMaker JumpStart

Amazon Web Services published a tutorial on its ML blog describing the deployment of the Qwen3-TTS-12Hz-1.7B-Base model from Amazon SageMaker JumpStart to a fully managed real-time inference endpoint. The model is publicly available and comes from the Qwen team at Alibaba Cloud. According to the article, it enables voice cloning from a short reference recording with a transcript, without needing to retrain the model – the model then delivers new text in the voice of the reference speaker.

The model covers 10 languages (Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish, Italian) and, according to the article, also supports cross-lingual cloning, where the voice is captured in one language and speech is generated in another while preserving the speaker's vocal identity. The Base variant used is designed for cloning from a few seconds of audio and can serve as a basis for further fine-tuning; it differs from the CustomVoice variant, which works with a fixed set of predefined voices. Besides the tested model, Qwen3-TTS-12Hz-1.7B-CustomVoice and Qwen3-ASR-1.7B are also available in JumpStart.

Deployment is done using the Amazon SageMaker Python SDK via the JumpStartModel object, with JumpStart providing a ready-made serving container without the need to write a custom inference handler. According to the article, the model produces 24 kHz audio output and internally runs in two phases (talker and code2wav) on a single GPU. Deploying the 1.7B model only requires an ml.g6.4xlarge instance with a single NVIDIA L4 GPU (24 GB), with the model weights taking up 3.66 GiB (talker) and 0.45 GiB (code2wav), the rest of the memory being KV cache. The endpoint includes automatic scaling and monitoring via Amazon CloudWatch. The rest of the configuration and monitoring metrics description is missing from the source article.