When integrating PlayHT: strip SSML yourself before sending text, the API will silently drop the tags. Pick the voice engine for the job, PlayHT2.0 for quality, Play3.0-mini for real-time speed. Map Unauthorized errors to a credential check and QuotaExceeded to a billing check in your error handling. For cloned voices, create the voice on the PlayHT site first and pass the full S3 URL voice ID, a short name will not resolve.

Context: Web (tts-wrapper docs, PlayHT engine): documents real integration limits. There is no SSML support, tags get stripped, so SSML markup in your text is silently ignored. No native word timing support. API rate limits apply. Voice engine choice matters: PlayHT2.0 gives higher quality with emotions and style guidance but slower processing, Play3.0-mini is faster for real-time use with fewer style options. Voice cloning needs extra setup, and a custom voice ID is the full S3 URL format like s3://voice-cloning-zero-shot/...