AWS Polly — Text-to-Speech

Amazon Polly converts text to lifelike spoken audio. With 100+ voices across 40+ languages and four voice engine tiers — including a Generative engine added in 2025 — it's the go-to TTS choice for AWS-native stacks. It trades voice variety and peak quality (ElevenLabs wins those) for cost-effectiveness, scale, and seamless AWS integration.

Four Voice Engines

Languages & Voices

• 100+ voices total across all engines
• 40+ languages and variants — including English (US, UK, AU, IN), Spanish (ES, US, MX), French (FR, CA), German, Japanese, Korean, Hindi, Brazilian Portuguese, Arabic, and more
• SSML support — full Speech Synthesis Markup Language for fine-grained control: pauses, emphasis, speaking rate, pitch, whisper effect, phoneme overrides
• Custom lexicons — define exactly how brand names or technical terms are pronounced

boto3 Usage

For short text, use synthesize_speech() — it returns an audio stream you write to a file:

import boto3

polly = boto3.client("polly", region_name="us-east-1")

response = polly.synthesize_speech(
    Text="Welcome to the AWS Bedrock platform. Let's get started.",
    OutputFormat="mp3",
    VoiceId="Joanna",
    Engine="neural",   # standard | neural | long-form | generative
)

# Write the audio stream to a file
with open("output.mp3", "wb") as f:
    f.write(response["AudioStream"].read())

For long text that exceeds the 3,000-character limit, use the async task API:

# For long documents — output written to S3
response = polly.start_speech_synthesis_task(
    Text=long_text,
    OutputFormat="mp3",
    VoiceId="Matthew",
    Engine="long-form",
    OutputS3BucketName="my-audio-bucket",
    OutputS3KeyPrefix="polly-output/",
)
task_id = response["SynthesisTask"]["TaskId"]
print(f"Task started: {task_id}")

For SSML input, set TextType="ssml" and wrap your text in <speak> tags. The Engine parameter selects which engine tier to use — not all voices are available in all engines.

Pricing (per 1 million characters)

1 million characters ≈ 8–10 hours of audio depending on speaking rate. New AWS customers also receive $200 in credits (from July 2025) applicable to Polly usage.

How Polly Compares

Integration with Bedrock

Polly is the TTS layer in many Bedrock-powered voice pipelines:

• Voice chatbot: Transcribe (audio → text) → Bedrock LLM (Claude, Nova, etc.) → Polly (text → audio response)
• Video auto-dubbing: Transcribe video audio → Translate text to target language → Bedrock refines translation → Polly synthesizes dubbed audio
• Document narration: Extract text from S3 documents → Polly Long-Form engine → MP3 audio output
• IVR systems: Dynamic responses from Bedrock LLMs, spoken by Polly in real-time via Connect

Amazon Connect (AWS's contact centre platform) has native Polly integration — customer-facing voice applications built on Connect automatically use Polly for TTS without additional configuration.

When to Use Polly

Checklist: Do You Understand This?

• Amazon Polly has four engine tiers: Standard ($4), Neural ($16), Long-Form ($100), Generative ($30) per 1M chars
• 100+ voices across 40+ languages; Generative engine has 31 voices in 20 locales (as of late 2025)
• Use synthesize_speech() for short text; start_speech_synthesis_task() for long documents via S3
• SSML support gives precise control over pauses, emphasis, speed, and pronunciation
• Pairs naturally with Transcribe + Bedrock for full voice pipeline (STT → LLM → TTS)
• ElevenLabs leads on quality; Polly leads on AWS integration and volume cost