qwen3-tts

What it does

Plans, invokes, reviews, and documents Qwen3-TTS speech generation decisions. It covers hosted DashScope/Model Studio routes and open-weight checkpoints, guiding model/voice selection, realtime vs non-realtime, voice design, voice cloning, language handling, articulation controls, and production QA for narration, characters, assistants, audiobooks, ads, localization, or voice-enabled media.

How it works

• It distinguishes between hosted DashScope and open-weight deployments, advising when to use each based on infrastructure, privacy, and control needs.
• It provides a decision table to pick routes (e.g., non-real-time built-in voices via qwen3-tts-flash, instruct models for expressive needs, voice design for custom voices, or voice cloning for consented speakers).
• For non-real-time HTTP/SSE, it specifies using text, voice, and language_type with streaming toggle via X-DashScope-SSE header, and enforces segmentation, downloading returned audio URLs, and maintaining a per-segment manifest with model/region/voice/language/instructions/request_id/usage/path data.
• For real-time WebSocket, it documents required SDKs, parameters (model, URL, voice, language_type, mode, output format, sample rate, speech rate, volume, pitch, bitrate, instructions, optimization), and modes (server_commit vs commit).
• For open-weight paths, it instructs defining a local service contract (text length, sample rate, file format, enrollment storage, logging, GPU memory, concurrency, moderation).
• It covers voices and pronunciation rules, including a list of built-in voices, language coverage, dialect considerations, and guidance on single-language vs mixed-language handling.
• It documents instruction control usage (instructions parameter, languages, token limits), and recommends treating instructions as performance direction with optional optimization.
• It describes voice design workflows (brief, candidate generation, audition, save/record-keeping, and quota considerations).

When to use it

• Use when selecting Qwen3-TTS models, voices, real-time vs non-real-time pathways, voice design or cloning, dialect or multilingual needs, pricing/privacy constraints, or production QA for voice media.
• Use when planning production routes, documenting endpoints, or auditing model-region configurations before production.

What it can touch

• Tools listed: claude-code, codex, copilot, cursor. It references endpoints, models, voices, regions, and instruction fields as part of decision and documentation workflows.

Caveats

• Open-weight deployment requires local inference engineering and QA; hosted DashScope retains installation-specific constraints like quotas and billing. Model availability and region endpoints are region-specific and may vary; verify region and model lists before production. Instruction scope and supported languages for instruct variants may differ between non-real-time guides and realtime SDK references. Voice availability, dialect support, and cloning/design capabilities depend on the current Qwen-TTS model list and may change over time.