AWS Brings Alibaba's Qwen3-TTS Voice Cloning Model to SageMaker Real-Time Endpoints

AWS has published a technical guide for deploying Alibaba's Qwen3-TTS-12Hz-1.7B-Base text-to-speech model as a fully managed real-time inference endpoint on Amazon SageMaker AI, enabling voice cloning from short reference audio clips without retraining.

What the model does

Qwen3-TTS is a text-to-speech model family developed by the Qwen team at Alibaba Cloud. It supports 10 languages: Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish, and Italian. The family uses the Qwen3-TTS-Tokenizer-12Hz speech tokenizer and supports streaming generation for low-latency interactive use.

The Base variant covered in this deployment, Qwen3-TTS-12Hz-1.7B-Base, performs voice cloning from a few seconds of reference audio plus its transcript. It captures a speaker's timbre, pitch, and cadence and applies those characteristics to new text, including across languages — a reference clip in one language can produce output in another while preserving vocal identity. This differs from the separate CustomVoice variant, which draws from a fixed set of predefined speakers rather than user-supplied reference audio. Both variants, along with Qwen3-ASR-1.7B, are available through Amazon SageMaker JumpStart.

Deployment details

According to AWS, the model can be deployed via the SageMaker Python SDK using a JumpStartModel object pointed at model ID huggingface-ttsvoiceclone-qwen3-tts-12hz-1-7b-base (version 1.0.1). The reference deployment runs on an ml.g6.4xlarge instance with a single NVIDIA L4 GPU (24 GB), which AWS states is sufficient for the 1.7B-parameter model. Output audio is generated at 24 kHz.

The serving container runs vLLM-Omni and processes requests through two stages sharing the same GPU: a "talker" stage that generates speech tokens from text and reference voice, and a "code2wav" stage that renders those tokens into a waveform. AWS notes that GPU memory allocation is the critical configuration variable — because both stages share one GPU via a single gpu_memory_utilization setting, AWS recommends a value of 0.45 (via the SM_VLLM_GPU_MEMORY_UTILIZATION environment variable) so neither stage triggers an out-of-memory error at startup.

Requests to the endpoint include target text, base64-encoded reference audio, and the reference audio's transcript. Responses return synthesized audio in WAV format. AWS also points to CloudWatch metrics for monitoring and right-sizing the endpoint post-deployment.

Use cases cited by AWS

AWS lists several target applications: content localization that preserves an original speaker's voice across languages, consistent brand voices for contact-center and virtual-assistant responses, instructor- or author-voiced e-learning and audiobook content, creative prototyping of dialogue and voiceovers before studio production, and real-time conversational AI agents combining streaming ASR with low-latency synthesis.

AWS frames the self-hosted approach as giving teams control over cost (compute-based rather than per-character API pricing), data residency (audio stays within the customer's AWS account), and the ability to fine-tune the model for domain-specific use.

Pricing for the underlying model is not disclosed beyond standard SageMaker instance costs; the ml.g6.4xlarge instance used in the walkthrough is billed at standard EC2 G6 rates through SageMaker, not per-token or per-character.

What this means

This is a deployment and infrastructure update, not a new model release — Qwen3-TTS itself was previously released by Alibaba's Qwen team and is not new to this announcement. The news here is that AWS has packaged it for turnkey deployment via SageMaker JumpStart, competing directly with commercial voice-cloning APIs from providers like ElevenLabs on a self-hosted, compute-billed basis. For teams already invested in AWS infrastructure with data-residency requirements, this lowers the barrier to running voice cloning without sending audio to a third-party API. The tradeoff is operational: teams take on GPU instance management, memory tuning, and scaling decisions that a managed API would otherwise abstract away.