PollyVoice
Amazon Polly text-to-speech for Blueprints and C++

Overview

PollyVoice gives your game a voice for any text known only at runtime: player names, procedural quests, AI-generated dialogue, accessibility narration, or simply thousands of lines you do not want to record. Requests run on background worker threads with automatic retries; every result arrives on the game thread.

• Synthesize Speech: text or SSML to a UPollyVoiceSoundWave (16-bit PCM, 8 or 16 kHz) that plays like any sound: Standard, Neural, Long-Form and Generative engines, language codes, lexicons.
• Synthesize With Speech Marks: the sound plus viseme, word, sentence and SSML marks; Get Speech Marks for timings only.
• Synthesize And Play 2D / At Location, Synthesize To File (MP3, Ogg Vorbis, PCM, WAV), Preload Phrases.
• UPollyVoiceLipSyncComponent: plays speech, drives viseme morph targets with smooth blending and fires On Viseme, On Word, On Sentence, On SSML Mark and On Finished.
• 17 pure SSML builder nodes: breaks, emphasis, prosody, max duration, say-as, phonemes, aliases, languages, marks, whisper, newscaster style and more.
• Memory + disk cache keyed by text and voice settings: a line is paid for once, then plays instantly, even after a restart.
• Voices, lexicons and asynchronous S3 synthesis tasks; Amazon Cognito identity pools, static keys or the default AWS chain.

Requirements

Installation

1. Install PollyVoice from Fab into your engine, or copy the PollyVoice folder into your project's Plugins folder.
2. Enable PollyVoice in Edit > Plugins and restart the editor.
3. Open Project Settings > Plugins > PollyVoice: set Region, Identity Pool Id and a Default Voice.
4. C++ only: add PublicDependencyModuleNames.Add("PollyVoice"); to your Build.cs.

Setting up AWS

1. Pick a region where Polly offers the voices you need (Neural and Generative voices are not in every region).
2. In Amazon Cognito create an identity pool. Enable guest access if players may hear speech before signing in.
3. Give the pool's player role (guest and/or authenticated) only what the game client needs — synthesis, voices and reading lexicons:

{
  "Version": "2012-10-17",
  "Statement": [
    {
      "Sid": "PollyVoicePlayers",
      "Effect": "Allow",
      "Action": ["polly:SynthesizeSpeech", "polly:DescribeVoices"],
      "Resource": "*"
    },
    {
      "Sid": "PollyVoiceReadLexicons",
      "Effect": "Allow",
      "Action": ["polly:GetLexicon", "polly:ListLexicons"],
      "Resource": "arn:aws:polly:us-east-1:123456789012:lexicon/*"
    }
  ]
}

To limit players to your own lexicons, replace lexicon/* with names such as lexicon/MyGame* (a request that names a lexicon is checked against that lexicon's ARN).

Admin and tool nodes

Put Lexicon, Delete Lexicon and the three synthesis-task nodes are meant for editor tools, build pipelines and servers. Give them a separate role or user, never the player role:

{
  "Version": "2012-10-17",
  "Statement": [
    { "Effect": "Allow", "Action": ["polly:PutLexicon", "polly:DeleteLexicon"],
      "Resource": "arn:aws:polly:us-east-1:123456789012:lexicon/MyGame*" },
    { "Effect": "Allow", "Action": ["polly:StartSpeechSynthesisTask", "polly:GetSpeechSynthesisTask", "polly:ListSpeechSynthesisTasks"],
      "Resource": "*" },
    { "Effect": "Allow", "Action": "s3:PutObject", "Resource": "arn:aws:s3:::my-voice-bucket/voice/*" },
    { "Effect": "Allow", "Action": "sns:Publish", "Resource": "arn:aws:sns:us-east-1:123456789012:voice-done" }
  ]
}

Polly writes task output with the caller's permissions, so the caller needs s3:PutObject on the output bucket (and sns:Publish when a topic is given).

Quick start

1. Say something

1. Add Synthesize And Play 2D to Begin Play. Text Welcome back, commander., Options empty (uses the Default Voice).
2. From On Failure print Error > Message. Done: the line plays; the second time it plays straight from the cache.

2. A talking character

1. Add a Poly Voice Lip Sync component to your character. In its details set Viseme Map from Make Default Viseme Map (or type your own morph target names).
2. Call Synthesize With Speech Marks with the line and Options.Voice = Matthew.
3. On Success: Play Speech on the component with Speech > Sound (leave Marks empty — they are stored in the sound).
4. Bind On Word to highlight subtitle words and On Finished to continue the dialogue.

3. Expressive lines with SSML

Build Make SSML Document( SSML Prosody( "Get down", Rate 120 ) + SSML Break(Strong) + SSML Amazon Effect( "they can hear us", Whispered ) ) and pass it as Text — Text Type Auto recognises the <speak> document.

4. Loading screen

Call Preload Phrases with every line of the level. During gameplay use Get Cached Sound (instant, no network) or Synthesize Speech (also instant for cached lines).

Lip sync

Polly reports a viseme (mouth shape) code with a start time for every sound of the speech: p t S T f k i r s u @ a e E o O sil (codes are case-sensitive). Build Timeline adds end times; Evaluate Morph Weights turns the timeline into morph-target weights at any time, blending each shape into the next over Blend Time. The lip-sync component does this every frame and applies the weights to the owner's first skeletal mesh (or the mesh given to Set Target Mesh).

• Sync Offset (default 0.05 s) moves the mouth slightly ahead of the audio to hide output latency.
• Use Build Viseme Curves to bake float curves (time → weight) for Sequencer, Timelines or an Animation Blueprint instead of the component.
• Lip sync without playing audio: Play Timeline, for example for audio from a synthesis task with marks from Get Speech Marks.
• Export Subtitles SRT writes SRT subtitles from the sentence (or word) marks.

SSML builder

Every builder node wraps a fragment in one tag, so builders chain. Send player-typed text through Escape Text For SSML first. Neural and Generative voices ignore some tags (emphasis, pitch, whisper); Polly rejects invalid SSML with InvalidSsmlException.

Cache

Every synthesized line is stored under a key made from the text, voice, engine, language, lexicons, sample rate and text type. The memory cache (default 32 MB, least recently used lines dropped first) makes repeats instant; the disk cache (default 256 MB in Saved/PollyVoice) keeps lines across sessions as <key>.wav + <key>.json (speech marks). Speech marks are cached per type, so asking for visemes after words only requests the visemes. Set Bypass Cache in the options for one-off lines.

Node reference

85 nodes. Call = executable, Pure = no exec pins, Latent = async node with output pins (each fires exactly once on the game thread; Cancel on its return value stops it).

Synthesis

Cache

Lip sync component (UPollyVoiceLipSyncComponent)

Lip sync and timing helpers

SSML builder

Sounds and audio files

Voices

Lexicons

Synthesis tasks (S3)

Config and identity

Credentials

C++ API

// Build.cs: PublicDependencyModuleNames.Add("PollyVoice");
#include "PollyVoiceLibrary.h"
#include "PollyVoiceSoundWave.h"

FPollyVoiceSynthesisOptions Options;
Options.Voice = TEXT("Amy");
UPollyVoiceLibrary::SynthesizeNative(TEXT("Hello there."), Options, /*bWithSpeechMarks*/ true,
    FPollyVoiceNativeDelegate::CreateLambda([this](const FPollyVoiceSpeech& Speech, const FPollyVoiceError& Error)
    {
        // game thread; keep Speech.Sound referenced (UPROPERTY) or play it now
        if (Speech.Sound) { LipSync->PlaySpeech(Speech.Sound, Speech.Marks); }
    }));

The latent node classes (UPollyVoiceSynthesizeAction and the others) can also be created from C++ through their static factory functions; bind the delegates, then call Activate().

Types

Settings

Project Settings > Plugins > PollyVoice (saved to DefaultGame.ini). Changes apply to the next request.

Engine notes

• UE 4.27: PollyVoice sounds play from in-memory PCM (like editor previews). Looping, attenuation, concurrency and sound classes work as usual.
• UE 5.8: PollyVoice sounds stream through a sound generator, one per playing instance, so the same sound can play several times at once.
• Sounds are created at runtime and are not assets: keep them in a variable (or a UPROPERTY in C++) as long as you want to replay them; the cache recreates them cheaply.
• Polly returns mono 16-bit PCM at 8 or 16 kHz for playable sounds; use Synthesize To File with MP3 / Ogg at 22050 or 24000 Hz for higher fidelity assets.
• Text limits per request: 3,000 billed characters (6,000 total with SSML tags). Use a synthesis task for longer text.

Troubleshooting

Credits

© 2026 Alpha XP — alphaxp.net · Support: support@alphaxp.net. Includes the AWS SDK for C++ and the AWS Common Runtime (Apache License 2.0). Amazon Polly is a service of Amazon Web Services. Built for Unreal Engine 4.27 and 5.8.