amazon-polly

What it does

Instructs an AI agent to plan and execute production text-to-speech tasks using Amazon Polly, selecting an engine and operation (SynthesizeSpeech, StartSpeechSynthesisTask, StartSpeechSynthesisStream), choosing compatible voices, and handling SSML, speech marks, and lexicons. It covers output formats, quotas, pricing, and workflow guidance for media like narration, audiobooks, and captions, plus QA and publishing considerations.

How it works

The skill directs the agent to:

• Decide between operations: SynthesizeSpeech (real-time), StartSpeechSynthesisTask (asynchronous to S3 with optional SNS), StartSpeechSynthesisStream (streaming over HTTP/2, generative-only with marks caveat).
• Then select an engine: standard, neural, long-form, or generative, with explicit notes on capabilities and limitations (e.g., speech marks availability varies by engine).
• Ensure the production use adheres to documented limits, formats, and region/voice compatibility, using DescribeVoices or current AWS pages for verification.
• Apply SSML and pronunciation controls as described, noting tag support and constraints per engine.
• Generate speech marks when timing artifacts are needed, requesting OutputFormat="json" and using the appropriate SpeechMarkTypes.
• Manage lexicons per region with constraints on counts and characters, and use them to fix pronunciations.
• Consider production heuristics: when to prefer Neural, Long-form, Generative, or Standard, and how to structure SSML and lexicons for stable reuse.
• Follow QA checks: validate audio, captions, timing, and potential mismatches after updates.

When to use it

Use when a production TTS job requires AWS-hosted voices with IAM governance, SSML control, and reusable audio assets. Prefer Neural for edited narration with marks and SSML control, Long-form for long chapters if voice constraints fit, Generative for conversational lines with QA review, and Standard for high-volume or telephony scenarios.

What it can touch

Polly engines, voices, SSML payloads, audio output formats (mp3, ogg_vorbis, ogg_opus, pcm, mulaw, alaw, json), speech marks (OutputFormat="json"), lexicons (up to 100 per region/account, up to 40,000 chars each), and S3 endpoints for asynchronous tasks. It references external AWS resources and pricing pages for limits and costs.

Caveats

Notes include that Generative speech marks are not available and certain engine capabilities vary (e.g., some SSML tags unsupported by specific engines), and that model updates may affect timing or output. Pricing, quotas, and region availability are as of the cited AWS documentation date and should be rechecked for current values. The workflow emphasizes production-level safeguards rather than voice cloning or impersonation.