Methodology
How We Test AI Voice Models
A transparent method for separating official specifications, website functionality, and first-hand evidence.
These labels clarify the origin of a statement. A feature in official documentation may not be available through this site, and a site feature does not by itself prove a model-wide result.
What We Test
We assess the workflow question at hand rather than relying on a universal quality score. Depending on the feature, that can include pronunciation, pacing, consistency, clarity, style control, and the behavior of authorized reference audio.
Controlled Inputs
When publishing a comparison or example set, we keep the source text and the relevant source material fixed while changing only the variable under review. Any exception is described alongside the evidence.
Text-to-Speech Tests
We use representative scripts and listen for intelligibility, punctuation handling, pacing, and repeatability. A sample is evidence for that particular input and configuration, not a claim that every voice or language will behave identically.
Voice Cloning Tests
We use only permitted reference audio. We do not publish a universal similarity percentage or promise identity retention, because output depends on the reference, language, script, model, and implementation.
Style Control
We compare instructions against the generated result only when both the prompt and playable output are available. Descriptions of qualitative differences are labeled as observations, not model-wide facts.
Multilingual Speech
Official language support is reported as an official specification. Site-specific availability and any first-hand examples are identified separately; neither is used to infer quality in every language.
Failure Cases & Limits
We document known limitations when there is evidence to support them. Where evidence is incomplete, we state the boundary rather than invent a failure rate or benchmark result.
Sources & Updates
Primary documentation is preferred for model capabilities. Pages are updated when the product implementation or verified source material changes; we do not set update dates automatically.
Sources & evidence
Primary sources
- Qwen3-TTS official model repository
Official model documentation · QwenLM
- Qwen3-TTS CustomVoice model card
Official model page · Qwen
See the evidence hub
The benchmark page links to the current playable examples and records what they can—and cannot—show.
View the Qwen3 TTS Benchmark