Aws Polly

Polly provides high-quality TTS with Neural voices (NTTS) across 60+ languages and 100+ voices. For agents generating voice audio: synchronous SynthesizeSpeech for short texts (returns audio stream directly), async StartSpeechSynthesisTask for longer content (writes MP3/OGG to S3). SSML support for pronunciation control, pauses, emphasis, and speaking rate adjustments. Neural voices sound significantly better than standard voices. Confidence is docs-derived.

AWS SDK (boto3): polly = boto3.client("polly", region_name="us-east-1"). response = polly.synthesize_speech(Text="Hello, world!", OutputFormat="mp3", VoiceId="Joanna", Engine="neural"). audio = response["AudioStream"].read(). Async (long text): polly.start_speech_synthesis_task(Text=long_text, OutputFormat="mp3", VoiceId="Joanna", Engine="neural", OutputS3BucketName=bucket, OutputS3KeyPrefix="polly/"). SSML: Text="<speak>Hello <emphasis>world</emphasis></speak>", TextType="ssml".

IAM auth: polly:SynthesizeSpeech, polly:StartSpeechSynthesisTask, polly:GetSpeechSynthesisTask. S3 write for async tasks. Managed identity for EC2/Lambda. HTTPS enforced.

TextLengthExceededException for texts >3000 characters (sync) — use async for longer text. InvalidSsmlException for malformed SSML. Async task states: scheduled → inProgress → completed. Polly SLA: 99.9%. Neural voices available only in specific regions — check Voice.SupportedEngines before deployment. Audio quality is excellent for NTTS voices.

docs.aws.amazon.com/polly covers voice list, SSML reference, sync vs async synthesis, and lexicon management. Getting started: polly client + synthesize_speech in under 5 minutes. Free tier: 5M characters/month for first 12 months (standard); 1M characters/month (neural). Community via AWS Forums.