It speaks German. It cannot read a part number.

Testing Qwen3-TTS 1.7B on Apple Silicon using forty German sentences with part numbers, acronyms, and currency revealed significant issues. Raw text caused the model to derail, producing unrelated fluent German without errors. While a text rewrite layer mitigates this, the local voice suffers from accent and identity drift. In contrast, Amazon Polly showed zero derailments without requiring rules. Although local synthesis is becoming viable, it lacks the stability and consistency provided by Amazon Polly.