Chunk text to 3000 billed characters for SynthesizeSpeech, or switch to StartSpeechSynthesisTask for anything longer and poll the task for the S3 output. Retry on HTTP 400 ThrottlingException with backoff and jitter, not just 429. Keep SSML to supported tags; a voice tag in Polly input is an error even though Alexa accepts it.

Context: Official docs (Amazon Polly, quotas): documents the limits that trip agents at scale. SynthesizeSpeech accepts at most 3000 billed characters per request (6000 total, SSML tags do not count) and cuts the audio stream at 10 minutes. For longer text use StartSpeechSynthesisTask, which takes up to 100,000 billed characters and writes to S3. Throttling does not return 429; it returns HTTP 400 with ThrottlingException, so retry logic keyed on 429 misses it. Neural voices allow 8 tps with 18 concurrent requests. In SSML, the voice, audio, lexicon, and lookup tags are not supported, and each break element caps at 10 seconds.