Speech recognition has improved substantially. It still degrades in ways that matter in real deployments, and the useful skill is designing an interaction that tolerates the degradation rather than assuming it away.
Accents and pronunciation
The largest practical issue.
The situation. Recognition systems are typically trained more heavily on standard pronunciation, so accuracy falls with strong regional accents and with non-native speakers.
The symptom. A person speaks and the robot mishears or fails entirely, requiring repetition. This produces irritation faster than any other failure mode.
How to reduce the impact. Always offer a touch alternative — display common questions on screen so they can be selected rather than spoken. This is the most effective and cheapest mitigation.
Test with real local speakers. Before purchase, have several people with different accents try it. This is a test no supplier can prepare for in advance.
Ask about training data. How much the system has been trained on the relevant language and accents, or whether it is an international product with the language added.
This is the clearest differentiator between products and the least discussed in marketing material.
Noise and environment
Not a language issue but it directly affects comprehension.
Real environments are loud. Open lobbies, restaurants, retail spaces, traffic noise from outside. Noise levels far exceed the closed office where products are designed and tested.
A practical threshold. Above a certain noise level every system performs poorly, regardless of product.
Mitigations. Place the robot away from noise sources. Use directional microphones. Favour screen interaction in loud locations. And lower expectations for the speech component at noisy positions.
Multiple people speaking. Visitors arrive in groups and talk over one another. The robot struggles to identify who is addressing it.
People speaking quietly. Many are reluctant to speak loudly to a robot in front of others. The robot cannot hear, they become more self-conscious, and they give up.
That last point suggests something practical: in busy public spaces, screen interaction is frequently better than voice, however much less impressive it appears.
Numbers, names and foreign words
A specific and easily fixed group of errors.
Reading monetary amounts. Large figures can be read several ways. Standardise the reading in the content rather than leaving it to the system.
Reading phone numbers. Digit by digit or in groups. Specify it.
Proper nouns. Company names, product names, place names — the robot may pronounce them wrongly. Check and specify pronunciation where the system allows.
Foreign words inside sentences. Very common in product names and technical terms, and often rendered oddly.
Abbreviations. Write them out in full so the robot reads them correctly.
Units of measurement. Square metres, kilograms, percentages — check how each is spoken.
An effective check: listen to the robot read all the important content once through before going live. An hour of work catches nearly all of this group.
Register and how it addresses people
This affects perception more than most people expect.
Choose a consistent register. Suited to your customers. What fits a café does not fit a clinic.
Do not attempt to guess age or status. Systems that adapt address based on appearance get it wrong sometimes, and getting it wrong is worse than being slightly formal.
Consistency across all content. Mixing registers between sections is a common error when several people write content.
Adjust by sector. A clinic, a café and a car showroom call for different tones. This is a brand decision rather than a technical one.
Shorter than you think. A listening person has less patience than a reading one. Three sentences, with detail on request.
Prepare the opening and closing carefully. They are the only two lines every visitor certainly hears, and they set the impression of the whole interaction.
Designing to depend less on speech
The most practical conclusion of the whole topic.
Always provide a touch path alongside the speech path. Anything achievable by speaking should be achievable by touching the screen. This is the single most important design principle for public environments.
Show what was heard. So the person can see whether it understood and correct it. Simple, and it removes a great deal of frustration.
Display answers on screen. Alongside speaking them. In noisy places people read instead of listening.
Make retrying easy. A clear repeat option, without forcing the person to wait through a wrong answer.
Keep required utterances short. Short phrases are recognised more accurately than long ones. Design the conversation as short steps rather than one open-ended question.
Accept that a share of users will not speak at all. And design so they can still use the robot fully. That is not a failure; it is the reality of public spaces.
Frequently asked questions
What is the largest practical speech issue?
Regional accents. Recognition systems are trained more heavily on standard pronunciation, and having to repeat yourself produces irritation faster than any other failure mode.
How do you test speech quality before buying?
Have several local speakers with different accents talk to the robot directly. This is a test no supplier can prepare for in advance and it gives a clearer answer than any specification.
Which error group is easiest to fix?
Numbers, proper nouns, foreign words and abbreviations. Listening to the robot read all important content once before going live catches nearly all of them in about an hour.
What is the most important design principle for public spaces?
Always provide a touch path alongside the speech path. Public environments are noisy and many people are reluctant to speak aloud to a robot, so everything achievable by voice must be achievable by touch.
More in Deployment and operations and Robots for real estate and showrooms.
