Docs / Providers
Alibaba Cloud Bailian (CosyVoice / Qwen-TTS)
In one sentence: Soniva's default brand. Natural-sounding Chinese, the most voices, pricing in the middle of the pack — when you don't know what to pick, this one won't steer you wrong.
In one sentence: Soniva's default brand. Natural-sounding Chinese, the most voices, pricing in the middle of the pack — when you don't know what to pick, this one won't steer you wrong.
What This Is
Alibaba Cloud Bailian (Model Studio) is Alibaba Cloud's large-model service platform. Its speech synthesis has two product lines:
- CosyVoice series — the flagship line, most mature Chinese performance, most voices.
- Qwen-TTS / Qwen-Audio-TTS series — the Tongyi Qianwen team's speech models, fewer voices but faster to update.
Soniva's default is this brand's CosyVoice v3-flash; it's the first thing a new user sees on opening TTS Settings.
Selectable Models & Pricing
| Model | Credits / 1,000 chars | Number of voices | Notes |
|---|---|---|---|
| cosyvoice-v3-flash | 16 | Most (~91) | Default tier. Best value; use this for the vast majority of scenarios |
| cosyvoice-v3-plus | 32 | Only 2 | Twice the price, yet very few voices — generally not recommended |
| cosyvoice-v2 | 32 | Moderate | Previous generation; voices aren't shared with v3 |
| cosyvoice-v1 | 32 | Moderate | Earliest generation |
| qwen-audio-3.0-tts-plus | 32 | Only 2 | Official newest generation; currently only 2 system voices |
| qwen-audio-3.0-tts-flash | 32 | Few | Same as above, faster tier |
| qwen3-tts-instruct-flash | 32 | Few | Supports describing tone via natural-language instruction |
| qwen3-tts-flash | 32 | Few | |
| qwen-tts | 32 | Few | Earliest Qwen speech model |
Why all Qwen series are marked 32: the Qwen series is billed per token upstream, not per character, so it doesn't share the same basis as the rate table. Until Soniva converts it, these models are priced conservatively at Alibaba Cloud's most expensive tier — meaning credits get deducted at no less than what you'll actually be charged. It'll be adjusted down once verified.
Price source: https://help.aliyun.com/zh/model-studio/model-pricing (verified 2026-08-17)
Which Model to Pick
Just use the default cosyvoice-v3-flash. It holds three firsts at once: the cheapest, the most voices, and the most natural Chinese.
Only two situations call for switching tiers:
- You've heard a specific voice elsewhere and it only exists under v2 / v1 — CosyVoice's four versions have completely independent voice sets, not different renderings of the same batch. v3-plus only has 2 voices, a completely different thing from v3-flash's 91.
- You want to try
qwen3-tts-instruct-flash's instruction control — describing the desired tone in natural language.
Voices
Alibaba Cloud doesn't expose a public voice-list interface, so Soniva uses a built-in static voice table, cross-checked line by line against the official per-model voice docs (CosyVoice's four versions and Qwen's three model generations are each independent).
The free-preview representative voice is Long Anyang (under cosyvoice-v3-flash).
About "Voice Cloning"
Alibaba Cloud also has cosyvoice-v3.5-flash / v3.5-plus / qwen3-tts-vd / qwen3-tts-vc
models. Their names look like newer versions, but they're voice-cloning / voice-design models,
not ordinary preset-voice models — selecting one leaves the official voice list itself empty;
you must first upload your own voice for cloning before you can synthesize with the cloned voice.
Soniva currently doesn't offer voice cloning, so these models aren't in the selectable list. They'll be re-connected once the cloning feature is done.
Bring Your Own Key (BYOK)
Alibaba Cloud is the only brand whose two key types use two different protocols — note this when filling in a key:
| Key prefix | Source | Protocol |
|---|---|---|
sk-... | DashScope master account key | Regular HTTP interface |
sk-ws-... | Bailian workspace key | WebSocket only |
Soniva auto-detects the prefix and switches to the corresponding protocol — no setup needed; both key types work once filled in.
Using your own key deducts no Soniva credits, and the per-reading cap also relaxes from 500 to 4,000 characters.
Related Documents
Was this article helpful?
Anonymous feedback is enabled when you allow analytics and the service is available. You can also report an issue by email.
The email includes this article. You can also copy the details and send them to soniva.cde@gmail.com