VoiceScribe — Documentation

Overview

VoiceScribe turns speech into text entirely on the local machine. No network calls, no API keys, no per-request cost, and nothing leaves the device.

Feed it audio — a recording, a microphone stream, or an imported file — and it returns transcribed text, either in one pass or segment by segment as the speaker talks.

Fully offline — the model ships inside your packaged build
Streaming or one-shot recognition
Multi-language, with optional translation to English
Selectable model size, from Tiny to Large, including quantised variants
Editor tool to download and manage models
Tunable decoding: threads, beam size, temperature, prompts, token limits

Privacy Because recognition runs locally, VoiceScribe suits projects with privacy or compliance constraints, and it keeps working with no internet connection.

Requirements

Requirement	Details
	Unreal Engine	4.27 and 5.8 (verified clean on both)
	Modules	`VoiceScribe` (Runtime), `VoiceScribeEditor` (Editor)
	CPU	x64 with AVX. Recognition is CPU-bound.
	Disk	Model dependent — roughly 75 MB (Tiny) to 3 GB (Large)
	Language	Blueprints and/or C++
	Platforms	Win64 — built and verified. The code carries no Win64-only restriction, but other targets have not been verified by us.

Installation

Copy the VoiceScribe folder into your project's Plugins/ directory.
C++ projects: regenerate project files and build. Blueprint-only projects: launch the editor and accept the build prompt.
Open Edit → Plugins and confirm VoiceScribe is enabled.
Go to Project Settings → VoiceScribe, choose a model size and language, then download the model. This is a one-off editor step.
From C++, add "VoiceScribe" to PublicDependencyModuleNames.

Pick the model before packaging The selected model is baked into the build. Start with Tiny or Base while prototyping — the larger models are far more accurate but much slower and much bigger.

Quick start

Transcribe a buffer of audio

Call Create Speech Recognizer and keep the result.
Bind On Recognized Text Segment and On Recognition Finished.
Call Start Speech Recognition.
Push PCM in with Process Audio Data as it becomes available.
Call Stop Speech Recognition when the speaker is done.

USpeechRecognizer* Recognizer = USpeechRecognizer::CreateSpeechRecognizer();

Recognizer->OnRecognizedTextSegment.AddDynamic(this, &AMyActor::HandleSegment);
Recognizer->OnRecognitionFinished.AddDynamic(this, &AMyActor::HandleFinished);

Recognizer->SetStreamingDefaults();
Recognizer->SetLanguage(ESpeechRecognizerLanguage::En);
Recognizer->StartSpeechRecognition();

// then, as audio arrives:
Recognizer->ProcessAudioData(PcmSamples, SampleRate, NumChannels, /*bLast=*/false);

Streaming versus one-shot Set Streaming Defaults favours latency and emits partial segments as the speaker talks. Set Non Streaming Defaults favours accuracy on a complete recording.

Recognition API

Type	Description
Create Speech Recognizer	Callable	Create a recognizer instance.
Start Speech Recognition	Callable	Begin a recognition session.
Process Audio Data	Callable	Push PCM samples in.
Stop Speech Recognition	Callable	End the session and flush.
Force Process Pending Audio Data	Callable	Process buffered audio immediately.
Clear Audio Data	Callable	Discard buffered audio.
Get Is Finished / Get Is Stopped / Get Is Stopping	Pure	Session state.
On Recognized Text Segment	Event	Fires per transcribed segment.
On Recognition Finished	Event	Fires when the session completes.
On Recognition Progress	Event	Percentage progress.
On Recognition Stopped	Event	Fires after an explicit stop.
On Recognition Error	Event	Fires with a message on failure.

Tuning recognition

Type	Description
Set Streaming Defaults	Callable	Preset tuned for low latency.
Set Non Streaming Defaults	Callable	Preset tuned for accuracy.
Set Language	Callable	Spoken language of the audio.
Set Translate To English	Callable	Translate the transcript into English.
Set Num Of Threads	Callable	CPU threads used for decoding.
Set Beam Size	Callable	Beam search width. Higher is slower and more accurate.
Set Step Size	Callable	Audio chunk size for streaming.
Set Audio Context Size	Callable	How much past audio to keep as context.
Set Max Tokens	Callable	Cap tokens per segment.
Set Initial Prompt	Callable	Bias decoding with expected vocabulary.
Set Temperature To Increase	Callable	Fallback temperature on low confidence.
Set Entropy Threshold	Callable	Threshold that triggers the temperature fallback.
Set Single Segment	Callable	Force one segment for the whole input.
Set No Context	Callable	Do not carry context between segments.
Set Suppress Blank	Callable	Suppress blank outputs.
Set Suppress Non Speech Tokens	Callable	Suppress non-speech markers.
Set Speed Up	Callable	Trade accuracy for speed.
Get / Set Recognition Parameters	Callable	Read or apply the whole parameter struct.
Get Streaming / Non Streaming Defaults	Pure	Inspect the presets.
Compute Levenshtein Similarity	Pure	Compare two strings — useful for voice commands.

Choosing a model

Set the model under Project Settings → VoiceScribe. Quantised variants (Q5_1, Q8_0) are considerably smaller and faster with a modest accuracy cost.

Rough size	Best for
Tiny / Tiny Q5_1 / Tiny Q8_0	~75 MB	Prototyping, short commands, mobile
Base / Base Q5_1	~140 MB	Good general default
Small / Small Q5_1 / Distil Small	~460 MB	Noticeably better accuracy
Medium	~1.5 GB	High accuracy, desktop only
Large	~3 GB	Best accuracy, slowest

Packaged size The model ships with your build. A Large model adds roughly 3 GB to the package — budget for it, or pick a quantised variant.

Configuration

Recognition runs on the CPU. No optional acceleration backends are bundled, so there is nothing to enable — the settings below are the only build-time knobs.

Switch	Default	Effect
	AVX baseline	on	CPU instruction baseline. Set via `MinCpuArchX64` on UE 5.3+ and `bUseAVX` on older engines. Lower it in the Build.cs if you must support pre-AVX hardware.
	PCHUsage	`NoSharedPCHs`	The module raises the CPU baseline, so it builds its own PCH rather than sharing the engine's.

Editor settings Model size, language and download management are under Project Settings → VoiceScribe, provided by the VoiceScribeEditor module.

Troubleshooting

Symptom	Cause and fix
	No text comes back	Confirm a model is downloaded in Project Settings, that the audio is 16-bit PCM at a sane sample rate, and that you called Start Speech Recognition before pushing data.
	Recognition is very slow	Use a smaller or quantised model, raise Set Num Of Threads, and lower Set Beam Size.
	Transcript is garbled	Usually the wrong language. Set it explicitly with Set Language rather than relying on detection, and give an Initial Prompt for domain vocabulary.
	Packaged build is enormous	The selected model is included in the package. Choose a quantised or smaller model.
	Crash or illegal instruction on old CPUs	The module targets an AVX baseline. On pre-AVX hardware, lower it in the Build.cs and rebuild.

Credits

All licence notices, including third-party components, ship at Source/ThirdParty/LICENSE-ThirdParty.txt.

VoiceScribeOffline speech recognition

Overview

Requirements

Installation

Quick start

Transcribe a buffer of audio

Recognition API

Tuning recognition

Choosing a model

Configuration

Troubleshooting

Credits

VoiceScribe
Offline speech recognition