What's New

AWS has made three Qwen3 speech AI models available in Amazon SageMaker JumpStart: Qwen3-TTS-12Hz-1.7B-CustomVoice, Qwen3-TTS-12Hz-1.7B-Base, and Qwen3-ASR-1.7B. These models cover the full spectrum of enterprise speech workloads, offering multilingual text-to-speech with customizable voice styles, rapid voice cloning, and automatic speech recognition across 52 languages and dialects. Customers can deploy any of these 1.7B-parameter foundation models with a few clicks directly from SageMaker Studio or via the SageMaker Python SDK.

How It Works

• Qwen3-TTS-12Hz-1.7B-CustomVoice generates speech from text using instruction-driven control, allowing developers to specify timbre, emotion, and prosody parameters at inference time across 10 supported languages.
• Qwen3-TTS-12Hz-1.7B-Base performs rapid voice cloning by accepting as little as 3 seconds of reference audio input, then synthesizing new speech that matches the target voice's characteristics — suitable as a fine-tuning base for domain-specific TTS.
• Qwen3-ASR-1.7B performs automatic speech recognition (ASR) across 52 languages and dialects, designed to maintain high accuracy in complex acoustic environments and supporting both streaming (real-time) and offline (batch) inference modes.
• All three models are deployed as SageMaker real-time inference endpoints, managed through SageMaker JumpStart's one-click deployment workflow or programmatically via the SageMaker Python SDK.
• The models run on AWS-managed infrastructure, meaning customers benefit from SageMaker's built-in auto-scaling, endpoint monitoring, and IAM-based access control without managing the underlying compute.

Why It's Important

• Reduced time-to-production for voice AI: Enterprises can go from zero to a deployed, production-grade speech model in minutes rather than weeks, eliminating the need to source, containerize, and optimize models independently.
• Broad multilingual coverage: With ASR supporting 52 languages/dialects and TTS supporting 10, organizations serving global or multilingual customer bases can address a wide range of localization requirements from a single platform.
• Voice cloning democratizes custom voice creation: The 3-second cloning capability in the Base TTS model dramatically lowers the barrier to creating branded or personalized voice experiences without extensive voice recording sessions.
• Instruction-driven speech control unlocks expressive applications: The ability to programmatically control emotion, prosody, and timbre opens new possibilities for dynamic content — such as audiobooks, virtual agents, and accessibility tools — that require nuanced, context-aware speech.
• Consolidation on AWS infrastructure: Running speech models alongside other AWS services (Connect, Lex, Transcribe, Bedrock) simplifies architecture, reduces data egress costs, and keeps sensitive audio data within a customer's AWS environment.

How It's Different

• Versus Amazon Transcribe / Polly: These are fully managed AWS services with fixed model capabilities; the Qwen3 models in JumpStart give developers direct access to the model weights and inference parameters, enabling fine-tuning, custom prompting, and deeper integration control.
• Versus self-hosted open-source speech models: Deploying Whisper or other OSS ASR models on EC2 requires manual container builds, scaling logic, and endpoint management; JumpStart abstracts all of that while still providing model-level flexibility.
• Rapid voice cloning at 3 seconds: Many commercial TTS solutions require minutes of reference audio for voice cloning; the Base model's 3-second requirement is significantly more practical for real-world, low-latency applications.
• Instruction-driven prosody control: Unlike most TTS APIs that offer limited SSML-style controls, the CustomVoice model accepts natural-language instructions for voice style, enabling richer and more intuitive control over speech output.
• 52-language ASR coverage: This exceeds the language coverage of many competing managed ASR services and is particularly strong for less-resourced languages and dialects often underserved by mainstream providers.
• 1.7B parameter efficiency: At 1.7B parameters, these models are designed to deliver strong performance at a lower inference cost compared to much larger speech foundation models, making them viable for high-throughput or cost-sensitive workloads.

When to Prefer It

• Customer-facing virtual assistants and IVR systems that require natural, emotionally expressive speech output in multiple languages — use Qwen3-TTS-12Hz-1.7B-CustomVoice for real-time, instruction-controlled voice generation.
• Branded voice or personalized audio experiences (e.g., audiobooks, e-learning, accessibility readers) where a specific voice identity must be replicated quickly — use Qwen3-TTS-12Hz-1.7B-Base for rapid voice cloning from short reference clips.
• Multilingual transcription pipelines for contact centers, media, or compliance use cases spanning many languages — use Qwen3-ASR-1.7B for its broad 52-language coverage and robustness in noisy acoustic environments.
• Real-time captioning and live event accessibility applications that require low-latency streaming ASR — Qwen3-ASR-1.7B's streaming mode makes it suitable for live transcription scenarios.
• Fine-tuning and domain adaptation of TTS for specialized vocabularies (medical, legal, technical) — Qwen3-TTS-12Hz-1.7B-Base serves as a flexible foundation model that can be further trained on domain-specific data via SageMaker Training.
• Prototyping and experimentation where teams want to evaluate state-of-the-art speech models quickly without infrastructure overhead — JumpStart's one-click deployment makes it ideal for rapid proof-of-concept development.

Availability

• General Availability: All three models (Qwen3-TTS-12Hz-1.7B-CustomVoice, Qwen3-TTS-12Hz-1.7B-Base, Qwen3-ASR-1.7B) are generally available in Amazon SageMaker JumpStart as of May 14, 2026.
• Access method: Models are accessible via the Models section of SageMaker Studio or programmatically through the SageMaker Python SDK.
• Regional availability: Specific supported AWS regions were not detailed in the announcement; customers should check the SageMaker JumpStart documentation or the AWS Regional Services table for the latest region availability.
• Pricing: Customers pay standard SageMaker real-time inference endpoint pricing based on the instance type selected; there are no additional JumpStart model licensing fees stated, though customers should verify third-party model terms for the Qwen3 models.
• Model source: Models are provided by Qwen (Alibaba Cloud's AI research team) and hosted within AWS infrastructure via JumpStart's foundation model catalog.
• Limitations: TTS models support 10 languages for speech synthesis; ASR supports 52 languages/dialects — workloads requiring languages outside these sets will need alternative solutions.