Qwen3-TTS Voice Cloning

Overview

Deployed Qwen3-TTS, a 1.7 billion parameter text-to-speech model, on serverless GPU as an additional voice cloning endpoint alongside F5-TTS. The model uses flash attention for efficient inference, producing high-quality voice synthesis with natural prosody and intonation.

Generated audio is stored on S3 for reliable delivery and integration with the AssetFlow pipeline. The deployment provides a second TTS engine option, giving the platform flexibility to choose the best voice quality for different content types.

Key Features

• 1.7 billion parameter model with natural prosody
• Flash attention for efficient GPU memory usage and faster inference
• Voice cloning from reference audio samples
• S3 storage integration for generated audio
• RunPod Serverless deployment with auto-scaling
• Custom Docker container with optimized dependencies

Architecture

Text Input → Tokenization → Qwen3-TTS (1.7B) → Audio Generation → S3 Upload
                                     ↓
                           Reference Audio → Speaker Embedding
                           Flash Attention   (Voice Cloning)

Deployment:
Docker Container → RunPod Serverless GPU → API Endpoint → AssetFlow

Text is tokenized and processed by the Qwen3-TTS model with flash attention enabled for memory-efficient inference. A speaker embedding derived from reference audio drives voice cloning. The generated audio uploads to S3 and is accessible via the AssetFlow pipeline for narration workflows.

Sample Output

Tech Stack