Skip to main content

Fine-tune speech recognition for your STT engine

Genesys Cloud lets you fine-tune how your speech-to-text (STT) engine listens to callers in voice bot flows. You can adjust when the engine decides a caller has finished speaking, how long it listens, how sensitive it is to noise, and which words to prioritize. Some settings apply to every STT engine. Other settings apply only to specific engines, which you configure via Architect’s Advanced Configuration setting. This article explains each setting, which engines support it, and how to approach tuning.

The following table describes each setting, where to configure it, and which STT engines support the feature.

SettingWhere to configureGoogle CloudMicrosoft AzureAmazon TranscribeGenesys NativeDeepgram FluxDeepgram Nova-3
Max Speech TimeoutBot flow settings User Input > Speech Recognition Parameters
Voice End of Input TimeoutAsk for Action in the bot flow
Speech Detection SensitivityBot Flow Settings > User Input > Speech Recognition Parameters
Advanced Configuration (engine-specific JSON)Bot Flow Settings > User Input > Speech Recognition Parameters
Note: Speech detection sensitivity is supported only by Nuance Recognizer as a Service, a legacy bring your own technology (BYOT) engine, and has no impact on the included engines listed above.

Configure speech recognition settings 

  1. Click Menu > Orchestration > Architect. Architect opens in a new tab.
  2. Click or hover over the Flows menu and select the desired flow type.
  3. Click the flow that you want to open.
  4. Under Settings, click User Input. 
  5. Click Speech Recognition Parameters. 
  6. Under Max Speech Timeout, set the maximum duration of the utterance to collect from the user.
  7. (Optional) Under Advanced Configuration, enter an engine-specific JSON configuration.
  8. Save and publish the flow. 

        For more information about User Input settings, see .

        Speech recognition settings applicable to all STT engines

        The following settings are supported by the STT engines that you use in Architect.

        Speech recognition settings that apply to specific STT engines

        The following sections describe speech recognition settings that are supported by only specific STT engines that you use in Architect.

         

        Tuning guidance

        Change one setting at a time and, before you publish, test with realistic call audio. The table below lists common symptoms and settings to try first. These recommendations are starting points, not guarantees, because your results depend on your callers, audio quality, and language.

        SymptomGuidance

        The bot interrupts callers who pause mid-answer.

        Increase Voice End of Input Timeout on the Ask for action. On Deepgram Flux, also raise eotThreshold or eotTimeoutMs. 

        The bot feels slow to respond after the caller finishes.

        Decrease Voice End of Input Timeout. On Deepgram Flux, lower eotThreshold or eotTimeoutMs. 

        Callers reading out long numbers (account or serial numbers) get cut off between digit groups.

        Increase Voice End of Input Timeout on the Ask for Slot action. 

        Brand names, product names, or jargon are misrecognized.

        With Genesys Native and AWS, use . For Deepgram, add them to keyterms (Flux) or keyterm (Nova-3). If using a list slot, add synonyms.

        Background noise triggers the bot unexpectedly.

        • On Genesys Native, increase threshold or min_speech_duration_ms.
        • On Microsoft Azure, set speechStartEventSensitivity to Low. 

        Callers continue speaking for too long.

        Decrease Max Speech Timeout. 

        A language behaves differently from the others in the same flow.

        Use languageOverrides to tune that language separately. 

        Quiet speech or short responses are not detected.

        On Genesys Native, decrease threshold. If very short utterances are not detected, consider decreasing min_speech_duration_ms.

        Best practices 

        • Start with the engine defaults. Tune only when you see a specific problem. 
        • Keep a record of what you change and why.
        • Test each change with representative call recordings. 
        • Validate your JSON before you save it. Invalid JSON, such as a trailing comma after the last item, can prevent the settings from being applied. 
        • If your organization has data residency or compliance requirements, review Deepgram’s regional endpoint documentation before you enable Deepgram Flux or Deepgram Nova-3 in production. For more information, see . 

        Additional resources