AivexaNewsSearch
AI news for builders and product teamsChecked every hour

Deploying real-time personalized speech with Qwen3-TTS on Amazon SageMaker AI

Collected Sep 30, 2026

AWS has described deploying the publicly available Qwen3-TTS-12Hz-1.7B-Base text-to-speech model from Amazon SageMaker JumpStart to a fully managed real-time inference endpoint.

The walkthrough uses the Amazon SageMaker Python SDK. It constructs a JumpStartModel with model version 1.0.1, deploys it with the accept_eula argument, and invokes the endpoint with the Amazon SageMaker runtime client. The container serves the model to generate 24 kHz audio. JumpStart supplies the model artifacts and a pre-built serving container, so no custom inference handler is written.

Qwen3-TTS is a text-to-speech model family developed by the Qwen team at Alibaba Cloud. It covers Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish, and Italian, uses the Qwen3-TTS-Tokenizer-12Hz speech tokenizer, and supports streaming generation. The Base variant performs voice cloning from a few seconds of user audio and can serve as a base for fine-tuning. It differs from the CustomVoice variant, which generates speech from a fixed set of predefined speakers. Qwen3-TTS-12Hz-1.7B-Base is available in JumpStart alongside Qwen3-TTS-12Hz-1.7B-CustomVoice and Qwen3-ASR-1.7B.

For voice cloning, the model takes a reference audio clip and its transcript, captures vocal characteristics such as timbre, pitch, and cadence, and applies them to new text without retraining. Cross-lingual cloning is supported: a reference in one language can generate speech in another. The article lists benefits including personalization at scale, multilingual reach, cost aligned with compute usage rather than per-character API pricing, and data staying within the AWS account and managed endpoint.

The walkthrough uses an ml.g6.4xlarge instance (1x NVIDIA L4 GPU, 24 GB). The model runs as two stages, talker and code2wav, on the same GPU, so the shared setting SM_VLLM_GPU_MEMORY_UTILIZATION is set to 0.45 for each stage. Reported startup values include 3.66 GiB for talker model weights, 0.45 GiB for code2wav weights, and a GPU KV cache size of 56,928 tokens. Requests must pass the custom attribute route=/v1/audio/speech to reach the text-to-speech handler, and use the OpenAI speech schema with task_type set to Base for cloning. Amazon CloudWatch metrics are used to monitor and right-size the endpoint; the article also covers deleting the endpoint, endpoint configuration, and model to avoid ongoing charges.

Read at AWS Machine Learning Blog

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt

Deploy the publicly available Qwen3-TTS-12Hz-1.7B-Base text-to-speech model from Amazon SageMaker JumpStart to a fully managed, real-time endpoint, and clone a voice from a short reference clip. Cross-lingual cloning preserves the speaker's identity across languages.