Deepgram, the speech recognition and voice AI company, has identified five categories of pronunciation error that it says are responsible for significant customer escalation costs in large-scale voice agent deployments.
The company estimates that mispronunciations of account numbers, policy IDs and transaction amounts translate to between $1.8 million and $2.16 million in preventable annual costs for platforms handling 500,000 or more monthly voice interactions, based on escalation costs of $3 to $4 per call and an affected call rate of 15 to 18%.
Deepgram's report lists the five error categories as homograph disambiguation, numeric and entity formatting, proper nouns, acronyms, and domain vocabulary.
Homograph disambiguation refers to words spelled identically but pronounced differently depending on context, such as "lead" or "read", and the report warns that accuracy on this category can drop 25 to 40 percentage points between controlled testing environments and live production systems.
The company gives the example of the string "2025", which may function as a year, a quantity, or a confirmation code, each requiring a different spoken rendering, and notes that "$45.99" should be read aloud as "forty-five dollars and ninety-nine cents."
Recommended fixes include entity-aware preprocessing, Speech Synthesis Markup Language (SSML) tags that give text-to-speech systems explicit pronunciation instructions, and centralised lexicons that define how domain-specific terms should be spoken.
For high-volume deployments, Deepgram recommends versioned Pronunciation Lexicon Specification (PLS) lexicons, automated pronunciation tests integrated into software delivery pipelines, and regression testing across more than 500 conversation paths to prevent errors from being reintroduced after updates.
The report specifies testing minimums: at least 20 numeric edge cases, at least 50 homograph pairs tested with contextual variations, and full coverage of domain-specific proper nouns.
Related reading
- End-to-end speech models cut voice latency by up to 70%, Deepgram says
- Phoneme Error Rate offers deeper insight into speech model performance, says McGillivray
- Deepgram triples concurrency limits for AI voice agents
Deepgram cited production data showing that addressing voice agent quality issues, including pronunciation accuracy, contributed to a 70% reduction in missed calls at organisations using Five9, the cloud contact centre platform.
The company is encouraging platform builders to test pronunciation handling using its Console product and $200 in free credits before making production commitments.
The recap
- Five TTS pronunciation error categories identified from production deployments
- Pronunciation failures cause $1.8-2.16 million annual preventable escalation costs
- Testing minimums: 20+ numeric, 50+ homograph pairs, 100% proper nouns
