Our Voice Agents are being deployed globally by enterprise customers. In English, conversations hold up really well. But it is different for other languages.

Today's voice models are remarkable, and we work closely with many of the teams building them. But model capabilities and training coverage vary across languages, and real-world calls expose gaps that benchmarks may not capture. " Supported" and " Production ready" measure two different things.

 

What "supported" actually means

When a model provider says it supports a language, it usually means the model can transcribe or speak it, scored on clean, read-aloud audio. A phone call is nothing like that. It is compressed audio, background noise, pauses and a booking code spelled one letter at a time.

In our own testing, one leading speech model scored about 4% WER (word error rate) on a public Japanese benchmark. On our test set, the same model's error rate was about 21%. Same model, same language, 5X worse.

 

Why quality drops outside English?

Few things go wrong, and they compound across call flow:

• The data is lopsided. Voice models learn from what exists, and most of what exists is English. Less data means less exposure to accents, noisy phone audio etc.
• Errors stack up. A voice agent is a chain: speech recognition, then reasoning, then speech output. One misheard word becomes a wrong intent, a wrong record, and a wrong answer read back with total confidence.
• Letters and numbers are hard. Spelling a name or a reference code is where non-English calls fall apart. In one French deployment, we analyzed more than 60,000 spelled letters. F and S alone caused 17% of all mix-ups. Callers may accept a read-back that sounds right even when the captured letters are incorrect, leaving the underlying error unresolved.
• People do not speak one language at a time. A German employee asking for printer help drops an English feature name into a German sentence. In one call, "Follow-me drucken" came back as gibberish. Product names, IT tools, and acronyms trip up recognition in every language that is not English.
• Voice interactions are personal. An agent that interrupts a caller who is carefully spelling a last name makes the interaction harder. And a voice that mispronounces your company name or a city loses trust in the first ten seconds.

What we are doing about it

We built the ServiceNow voice platform on a simple principle: a language is not done when it is on the list. It is done when callers can finish what they called to do.

Here is what that looks like in the product today which is configurable on Assistant Designer for Voice:

• 29 languages and regional variants , from Canadian French to Mexican Spanish to Hindi, each with its own voice persona and welcome message.
• Caller language selection by keypad or by voice, so the agent never has to guess.
• Key term dictionary. Teach speech recognition to recognize product names, acronyms, and internal tools as your callers actually say them, per language, so it stops hearing a common word where someone said a brand name.
• Pronunciation dictionary. Control how the agent says the names, places, and product terms that matter most, with phonetic spelling or phonemes, per language.
• Adjustable noise cancellation , because an airport is not as quiet as an office.

Behind the scenes, we evaluate voice models language by language and pick the best one for the job, instead of betting everything on one provider.

The questions to ask instead

Do not ask how many languages a voice agent supports. Ask:

1. What is the task completion rate in your language of focus, on phone audio?
2. Can I teach it my vocabulary?
3. Can I control how it pronounces my brand?
4. What happens when a caller spells something?

Multilingual Voice AI should ultimately be measured by outcomes, not coverage lists. The goal is not simply for the agent to recognize or speak a language. It is for callers to complete their task accurately, naturally, and with confidence - just as they would expect when speaking with a human agent.