benchr
Plus is not a general “best voice” candidate. It is a narrower procurement decision: can one of Alibaba's currently listed preset voices speak the exact material, in the exact language, with a controllable style and an acceptable delivery path? If the answer is no, the model's broader quality is irrelevant.
The voice list is part of the API contract
Alibaba's current voice list ties each system voice to a specific model and documents Mandarin and English for the Plus voices shown there. The page warns that voices cannot be mixed across models and that an incompatible language can fail. That makes voice selection an integration input, not a cosmetic choice for the design team.
Do not infer Arabic support from the Qwen brand or from another Alibaba speech product. If Arabic is mandatory, the checked Plus documentation does not establish it. Keep the requirement unresolved until the provider documents the exact voice-language combination or a controlled account test proves it under your commercial terms.
Instruction control needs conflict tests
Plus accepts natural-language direction for delivery such as tone, pace, emotion, and timbre. A polished demo usually asks for one thing at a time; production scripts contain collisions. Test a calm safety notice with urgent wording, a fast disclaimer that must stay intelligible, names beside serial numbers, and a paragraph that changes from neutral explanation to a firm call to action.
Also test the constraints around the text itself. The listed voices do not document SSML, word timestamps, or pronunciation hot fixes. If your workflow depends on phoneme-level repair, synchronized highlighting, or exact cue points, treat that as a stop condition rather than assuming a hidden feature.
A reproducible workload for Plus
Build a fixed script pack from real product copy and keep the text unchanged across candidates. Include a short notification, a long help-center paragraph, person and place names, dates and identifiers, an abbreviation-heavy passage, punctuation that must not be spoken, and one style transition. Run every script with the same named voice, region, audio format, and instruction.
| Decision gate | What to run | Pass condition |
|---|---|---|
| Voice and language | Every production language with the exact listed voice | No unsupported pairing, substitutions, or broken names |
| Style control | Conflicting pace, tone, and emotion directions | Meaning stays clear and instructions do not distort wording |
| Streaming path | Long text, reconnect, cancellation, and repeated sessions | No duplicated or missing speech after state changes |
| Operational trace | Log model ID, voice, instruction, region, and output artifact | A reviewer can reproduce the approved take |
| Commercial fit | Representative monthly text and retry volume | Alibaba supplies current units, limits, and terms |
Record pronunciation defects by span, not with a single quality score. Log the input text, instruction, voice, request region, reconnect behavior, and the accepted audio file. A preference vote can tell you which take sounds nicer; it cannot tell an engineer why a release failed or whether the same output can be reproduced.
Operations: WebSocket state and disclosure
The model documentation places Plus on a WebSocket path. Test the full connection lifecycle: start, append text, finish, cancel, reconnect, and retry after a partial output. Make sure a retry cannot play both the abandoned and replacement audio. If the application must disclose synthetic media, the API documentation also describes an AIGC watermark option; decide whether it is enabled, preserved through transcoding, and recorded in the asset log.
When to avoid this model
Avoid Plus when the product requires a cloned or newly designed identity, Arabic support is non-negotiable but not documented, or the experience depends on SSML, word timestamps, or live pronunciation replacement. Also pause procurement while price, limits, retention, and regional processing terms are still unknown. Flash is the adjacent Qwen option when authorized cloning and lower latency matter, but it introduces a different governance problem rather than merely offering a faster Plus.
| Field | Official record | Decision effect |
|---|---|---|
| API model ID | qwen-audio-3.0-tts-plus | Pin this identifier with the chosen voice |
| Voice approach | Built-in voices; no cloning or voice design | Approve from the published inventory |
| Transport | WebSocket | Test session and retry state |
| Public price / token limit | Not published in checked docs | Do not finalize a budget |