Project Settings¶

All Aboleth STT settings are configured in Project Settings > Plugins > Aboleth Speech-to-Text. These map to DefaultGame.ini under the section [/Script/AbolethSTT.WhisperDeveloperSettings].

Runtime changes

Many of these settings can also be changed at runtime without reloading the model. See Runtime Settings API for the full list of setters and getters.

Model¶

Setting	Type	Default	Description
`ModelFilename`	String	`ggml-large-v3-turbo-q5_0.bin`	Active Whisper model filename. Resolved relative to the `Plugins/AbolethSTT/Models/` directory.
`VADModelFilename`	String	`ggml-silero-v6.2.0.bin`	Silero VAD model filename. Read-only — ships with the plugin.

Note

Changing ModelFilename requires reloading the STT system (LoadSTTSystem). Use the Model Downloader to fetch additional models.

Silero VAD¶

These settings map directly to whisper_vad_params in the whisper.cpp backend. They control when speech is detected and how segment boundaries are drawn.

Setting	Type	Default	Range	Description
`threshold`	Float	`0.5`	0.0 -- 1.0	Speech probability threshold. Higher values require more confidence before triggering onset.
`min_speech_duration_ms`	Int	`250`	0 -- 5000	Minimum consecutive milliseconds above threshold before speech onset is confirmed. Prevents transient noise from triggering detection.
`min_silence_duration_ms`	Int	`100`	0 -- 5000	Minimum consecutive milliseconds below threshold before speech offset is confirmed. Higher values tolerate longer pauses mid-sentence.
`max_speech_duration_s`	Float	`0.0`	0.0 -- 3600.0	Maximum duration of a single speech segment before forcing a new one. `0.0` disables the limit.
`speech_pad_ms`	Int	`30`	0 -- 1000	Padding added before and after detected speech. The onset reach-back captures audio slightly before the VAD triggered.
`samples_overlap`	Float	`0.1`	0.0 -- 2.0	Overlap in seconds when copying speech audio between segments. Ensures no audio is lost at segment boundaries.

Tuning VAD

See the Tuning Guide for practical advice on adjusting these values for different environments and use cases.

Performance¶

Setting	Type	Default	Description
`NumThreads`	Int	`4`	Number of CPU threads used by whisper.cpp. Only affects CPU-bound work — GPU inference is not thread-limited.
`GPUBackend`	Enum	`Auto`	GPU acceleration backend. `Auto` selects CUDA if available, then Vulkan, then CPU. Options: `Auto`, `CUDA`, `Vulkan`.
`bFlashAttention`	Bool	`true`	Enables Flash Attention for faster encoder inference. Recommended on for all GPU backends.
`bDynamicAudioContext`	Bool	`false`	Scales the encoder window to match actual audio length instead of always processing 30 seconds. Reduces latency on short utterances.
`MinAudioContext`	Int	`512`	Floor value for dynamic audio context. Must be a power of 2. Prevents the encoder window from becoming too small on very short audio.

MinAudioContext

Setting MinAudioContext below 512 may cause artifacts with Whisper models. The default of 512 is recommended for all supported models.

Language¶

Setting	Type	Default	Description
`Language`	String	`"auto"`	Language code for transcription (e.g. `"en"`, `"es"`, `"ja"`). `"auto"` enables automatic language detection per utterance.
`bTranslateToEnglish`	Bool	`false`	When enabled, translates non-English speech to English. Uses Whisper's built-in translation capability.

Audio¶

Setting	Type	Default	Range	Description
`InputGainDb`	Float	`0.0`	0.0 -- 20.0	Software microphone gain in decibels. `0` is passthrough (no amplification). `6` dB is approximately 2x gain. `12` dB is approximately 4x gain.
`TargetSampleRate`	Int	`16000`	--	Target sample rate for Whisper inference. Read-only — Whisper requires 16kHz mono audio.

Tip

If your microphone input is too quiet, increase InputGainDb before adjusting VAD thresholds. A gain of 6 dB doubles the signal amplitude and often resolves low-sensitivity issues without changing detection behavior.

Streaming¶

Setting	Type	Default	Range	Description
`bEnableStreaming`	Bool	`true`	--	Enables Local Agreement streaming. When on, partial transcription results are emitted while the user is still speaking.
`StreamingChunkIntervalMs`	Int	`1000`	50 -- 2000	Interval in milliseconds between streaming inference passes. Lower values give faster feedback but increase GPU load.
`BufferTrimmingSec`	Int	`6`	3 -- 30	When the accumulated audio buffer exceeds this length in seconds, earlier audio is trimmed. Keeps memory and inference time bounded during long utterances.

Beam Search¶

Beam search is disabled by default. It improves accuracy on the final transcription pass at the cost of additional compute.

Setting	Type	Default	Range	Description
`bEnableBeamSearch`	Bool	`false`	--	Enables beam search decoding on the final (post-utterance) inference pass.
`BeamSize`	Int	`5`	1 -- 10	Number of beams to maintain during search. Higher values explore more hypotheses.
`LengthPenalty`	Float	`-1.0`	-1.0 -- 3.0	Penalty applied to longer sequences. `-1.0` disables the penalty. Values above 1.0 favor longer transcriptions.
`bBeamSearchDuringStreaming`	Bool	`false`	--	Enables beam search during streaming passes. Significantly increases streaming latency.
`bEnableBeamSearchGate`	Bool	`true`	--	Enables adaptive beam dropout. Beams that fall too far behind the leader are pruned early to save compute.
`BeamSearchGateMs`	Int	`250`	50 -- 2000	Time threshold in milliseconds for the beam search gate. Beams not contributing within this window are dropped.

Note

Enabling beam search during streaming (bBeamSearchDuringStreaming) is generally not recommended. The latency cost outweighs the accuracy gain for real-time use cases. Reserve beam search for the final pass where accuracy matters most.

Debug¶

Setting	Type	Default	Description
`bEnableDebugLogging`	Bool	`false`	Logs per-utterance diagnostics including timing, segment lengths, and VAD decisions. Output goes to the `LogAbolethSTT` log category.
`bEnableProbabilityLogging`	Bool	`false`	Writes per-frame speech probability values to a CSV file for offline analysis. Useful for tuning VAD thresholds against specific audio environments.

Info

For more on debugging tools, see Debug & Profiling.