Real-time Voice Cloning with Qwen3‑TTS on SageMaker
AWS details how to deploy the Qwen3‑TTS‑12Hz‑1.7B‑Base model on SageMaker JumpStart for real‑time, personalized speech synthesis using voice cloning.
Original source published: September 25, 2026
AWS announced that the publicly available Qwen3‑TTS‑12Hz‑1.7B‑Base text‑to‑speech model can be deployed from SageMaker JumpStart to a fully managed real‑time inference endpoint. Developed by Alibaba Cloud’s Qwen team, the model supports ten languages, streaming generation, and voice cloning from a few seconds of reference audio without retraining.
The deployment uses the SageMaker Python SDK to create a JumpStartModel, selects a ml.g6.4xlarge GPU instance, and configures GPU memory utilization at 0.45 so that the talker and code2wav stages share a 24 GB GPU. Once the endpoint is InService, clients send an HTTP request with target text, a base64‑encoded reference clip and its transcript, and receive 24 kHz WAV audio.
The blog cites use cases such as content localization, e‑learning, and real‑time conversational AI, emphasizing that audio data stays within the AWS account and that costs align with compute usage.