Introducing Amazon Polly Bidirectional Streaming: Real-time speech synthesis for conversational AI

Amazon Web Services has launched a significant upgrade to its text-to-speech service with the Bidirectional Streaming API for Amazon Polly. This new feature fundamentally changes how conversational AI applications handle speech synthesis by enabling true duplex communication over HTTP/2. Unlike traditional TTS APIs that require complete text before synthesis begins, this API allows developers to stream text incrementally to Amazon Polly while simultaneously receiving synthesized audio bytes in real-time. The architecture eliminates the need to wait for full LLM responses before starting speech synthesis, dramatically reducing latency in voice-enabled AI applications.

Performance benchmarks reveal substantial improvements: processing time decreased by 39% (from 115 seconds to 70 seconds) for a 970-word text sample, while API calls dropped from 27 to just 1. The new StartSpeechSynthesisStream API supports flush configurations for immediate synthesis of buffered text and maintains a single persistent connection for both sending text and receiving audio. This eliminates the complex server-side logic previously required for low-latency implementations, where developers had to implement text separation, make multiple parallel API calls, and reassemble audio streams.

The technology specifically targets conversational AI applications powered by large language models like GPT-4, Claude, or Llama, where text generation happens incrementally at approximately 30ms per word. By allowing synthesis to begin as soon as the first tokens arrive, Amazon Polly's bidirectional streaming enables more natural, responsive voice interactions in virtual assistants, customer service bots, and interactive AI applications. The service uses Amazon's Generative engine with voices like Matthew and supports MP3 output at 24kHz, maintaining audio quality while dramatically improving responsiveness.