Configure speech recognition
Speech recognition converts caller audio into text for a Modular agent. It affects word accuracy, language switching, and when Wirevox decides that the caller has finished a turn.
Realtime agents do not use this selector. Their active provider owns speech recognition and turn detection inside the Realtime session.
Available Modular options
| Editor option | Best fit | Important behavior |
|---|---|---|
| Deepgram Nova-3 | General single-language calls | Uses the configured primary language and supports recognition key terms. |
| Deepgram Flux Multilingual | English/French or code-switching calls | Receives primary and secondary language hints and supports key terms. |
| Gladia | Multilingual and code-switching evaluation | Receives both configured languages and enables code switching when a secondary language exists. |
| Cartesia Ink 2 | English conversational turn detection | English only in the current editor; uses semantic turn detection and supports key terms. |
| Auto | Let Wirevox choose the configured default | Resolves to Flux when a secondary language is present and Nova-3 otherwise. |
OpenAI Whisper is displayed as coming soon and cannot currently be selected.
When you add a secondary language, Wirevox selects Flux unless Gladia is already active. Cartesia is unavailable for non-English or bilingual configuration.
Tune patience
The Low, Medium, and High patience profiles change how long a Modular agent waits for evidence that the caller has finished.
- Low responds faster but can cut off callers who pause while thinking.
- Medium is the balanced starting point.
- High waits longer and suits deliberate speakers or workflows with long identifiers, but increases response latency.
The exact provider parameters differ. Wirevox maps the selected profile to the active provider's endpointing or semantic turn controls. Evaluate the resulting conversation instead of comparing the profile names as if they were identical provider settings.
Add recognition key terms
Use key terms for words the recognizer regularly mishears: company names, product names, local place names, acronyms, and specialist terminology.
Enter terms separated by commas or new lines. Wirevox accepts up to 100 parsed terms, with up to 50 characters per term. The editor limits the raw field to 2,000 characters.
Good terms are exact and selective:
Wirevox
LG Home Comfort
Etobicoke
SKU AB-410
Do not paste a full glossary or paragraphs into this field. Too many weak hints can reduce accuracy. Recognition hints help the transcript; they do not control how the voice pronounces a word. Use the pronunciation dictionary for spoken output.
Choose with a controlled evaluation
- Build a set of short recordings or live scripts that reflect real callers.
- Include noise, accents, interruptions, names, dates, addresses, and domain terms.
- Keep the model, prompt, voice, and test script fixed.
- Compare transcript accuracy, premature turn endings, response delay, and action-field accuracy.
- Change the provider or patience profile one at a time.
- Retest the winning configuration over an actual phone connection.
The best transcript is not enough if the agent consistently responds before a caller finishes. Measure recognition and turn timing together.