
A voice tool only helps the whole team if it stays usable for those who cannot, or would rather not, talk to it. Four guarantees make the difference: everything you can do by voice you can also do on screen; the tool never judges the way people speak; it shows what it understood and is easy to correct; and it has been tested with the people concerned, not just with management. Without them, voice helps the most confident speakers and leaves everyone else behind.
Thursday, 10:15 am, during prep
Héloïse runs a forty-cover brasserie. Since Monday she has been trying a voice assistant for everyday questions: the next task, the opening checklist, how to clean the slicer. For her, it works first time.
For Kofi, her sous-chef, it is a different story. Kofi has worn two hearing aids since an ear infection went untreated ten years ago. He cannot wear an earpiece on top of them. A spoken answer from the speaker gets lost under the extraction hood. In the end he picks up the tablet, reads the sheet and waves to a commis to say he has what he needs. The screen helped him, because it gave the same answer as voice.
On the same team, a young commis stutters. He has never tried the assistant and has not told anyone why. And the dishwasher, who learned French in her forties, had to repeat her question three times before giving up. (Fictional people and restaurant.)
Héloïse thought she had equipped her team; in fact, the tool only worked for people who speak like her.
Why a voice tool can exclude people without meaning to
Not every accent is understood equally well
Speech recognition systems learn from recordings. Voices that are rare in those recordings are recognised less well. In 2020, a Stanford team (Koenecke et al., published in PNAS) ran interviews through five commercial speech recognition systems from major US tech companies. The average word error rate was 35% for Black American speakers and 19% for white speakers. The authors found the same gap on identical phrases, which pointed to pronunciation and rhythm rather than vocabulary.
The study is about American English and says nothing about any particular tool in another language. It still shows that a voice tool is never neutral by default, and its weak spots land on the people who already struggle most to be heard.
Stuttering and pauses
Someone who stutters may repeat a syllable, get stuck on a word or pause in the middle of a sentence. Many voice tools treat that pause as the end of the question and answer half a request. Some ask the person to repeat. For someone who stutters, "Sorry, I didn't catch that, could you say it again?" in front of three colleagues is exactly the moment to avoid.
A tired voice
Kitchen staff strain their voices all day. Hoarseness at the end of a shift, a few days of dysphonia after a cold, a lost voice: speech becomes quieter, rougher, harder to recognise. On the day someone's voice gives out, the tool should be just as useful as the day before.
Deaf and hard of hearing employees
A spoken answer is useless to someone who cannot hear it. An earpiece does not always fit alongside a hearing aid. A speaker next to a roaring hood is no better. For these employees, the answer has to exist as text, at the same moment, with the same content.
Shyness
Talking to a machine in front of colleagues is not natural for everyone. Some people will never want to, and that is their right. A tool that only works well out loud turns a personal preference into a professional disadvantage.
What a voice tool must guarantee
- Parity between screen and voice. Anything you can ask or do by voice must be possible on screen, with the same permissions, the same answers and the same actions. Voice stays an option, never a condition of access.
- A written answer alongside the spoken one. What the tool says also appears on screen. Anyone who missed it can read it.
- No judgement on how people speak. No diction score, no hesitation count, no comparison between employees: the tool understands the question without assessing the person asking it.
- Time to finish a sentence. A pause should not cut a request short. And when the tool has not understood, it should offer to switch to the screen rather than ask for endless repeats.
- Show what was understood before acting. If the tool prepares an action, it summarises it and the person confirms or corrects. A recognition error then becomes a line to fix, not an action done wrong.
- Correct without starting over. People should be able to fix a misheard word on screen without saying everything again.
- Testing with the people concerned. A tool tried only by managers has been tested on the easiest voices. Trials should include a range of accents, tired voices, and people who stutter or are hard of hearing, if they agree to take part.
What the law says about adjusting the job
In France, article L5213-6 of the Labour Code sets a clear duty towards recognised disabled workers. The employer must take "appropriate measures", based on the needs of a concrete situation, so that they can get a job, keep it, do it and progress in it. The article adds that the employer must make sure that the software installed on a disabled person's workstation and needed for their work is accessible.
The measures are due unless their cost is disproportionate, taking into account the public funding available to offset it (article L5213-10). And refusing to take them can amount to discrimination under article L1133-3. The text is available on Légifrance, the official French legal database. Outside France, check your local rules: the UK, for example, has its own duty to make reasonable adjustments.
In practice: if you require a voice tool for a task, you must be able to offer another route to a hard of hearing employee who needs one. A tool that already offers full screen and voice parity makes that easy. Adjustments are worked out with the person, the occupational health service and, where there is one, the disability lead. They do not need to be recorded in a management tool: disability is sensitive data.
Situation, obstacle, guarantee
| Situation | What goes wrong with voice only | What the tool must offer |
|---|---|---|
| Under-represented accent | Questions misunderstood, repeated | Trials with real voices, switch to screen without losing anything |
| Stuttering | Sentence cut at the first pause, asked to repeat in public | Time to speak, correction on screen, no endless "say that again" |
| Hoarse voice at the end of a shift | Worse recognition on the day you are tired | The same service on screen, no detour |
| Employee with hearing aids | No earpiece possible, speaker drowned out by the hood | Text answer every time, same actions on screen |
| Deaf employee | No information through sound | Full screen and voice parity, visible alerts |
| Shyness, reluctance to talk to a machine | Disadvantage if the screen is second best | Free choice, with no consequence or comment |
| Recognition error on an action | Action carried out wrongly | Summary before confirmation, fix a single word |
Common mistakes
- Testing the tool with management and concluding that it works. Office voices are not dish pit voices.
- Making the screen a cut-down version. If some actions only work by voice, the tool has created two classes of employee.
- Asking in public who "struggles with voice". Nobody should have to explain to the team why they prefer to read.
- Tracking who uses voice and who does not. A preference says nothing about performance.
- Recording a disability or a speech difficulty in the tool. That is sensitive data, and only the adjustment matters.
- Believing a published recognition rate says anything about your team. Only your own trials, with your voices and your noise, will tell you.
What to measure
- The share of voice actions that can also be done on screen. The target is 100%, and you can check it in an hour, feature by feature.
- Misunderstood questions during the trial, logged by time of service and by station, never by person.
- One anonymous question to the team after two weeks: "Can you do everything you need without speaking to the tool?" A single "no" is worth following up.
Where Tsuno San comes in
Tsuno San is the intelligent voice assistant for frontline teams, connected to the procedures, tasks, checklists, routine checks and site information in Tsuno. It is the same San on screen and by voice: anything done by speaking can also be done on screen, with the same permissions as the person using it. San Voice is available as a demo; San Audio, the hands-free experience, is under exploration and testing (see where San stands).
For actions, San works in two steps: it prepares and summarises, the person confirms. Nothing runs without confirmation, and it is logged. That is also what makes a misunderstanding recoverable. San judges no one, analyses neither emotions nor the way people speak, and Tsuno refuses to record a disability or a health condition against an employee.
On languages, the Tsuno interface is available in French and English. For San, we are working on French, English, Arabic and Vietnamese. We do not publish recognition rates: we would rather you judge for yourself in a demo, with your own team's voices, and we want to widen our own trials to the people this article describes. The data rules are on the Trust and data page.
Further reading
This article builds on what voice changes for people who work on their feet and non-native speakers in the kitchen: understanding without shame. To plan your trials, see how to test a voice assistant in a kitchen, and to choose the right channel for the moment, San on screen, San by voice: when to use which. On what a tool must never infer from a voice, read AI and emotions at work: what has been banned since 2025. The pillar page AI, voice and privacy at work covers fair access alongside the other trust questions. On the team side, see onboarding a multilingual team: a manager's guide and mixed languages in the kitchen.
Key takeaways
Judge a voice tool by the person who finds it hardest to use, not by the one who uses it best. Insist that everything stays possible without speaking, test with your team's real voices, and let everyone choose to read rather than talk.
Frequently asked questions
Does speech recognition understand some accents less well than others?
Yes, and it is documented. A study published in PNAS in 2020 measured, across five commercial speech recognition systems, a word error rate almost twice as high for Black American speakers as for white speakers. A voice tool should therefore be tried with the real voices of your team.
Can an employee who stutters use a voice assistant?
They should be able to if they want to, with a tool that does not cut them off after a pause and that shows what it understood before acting. Above all, they must be able to do everything on screen without ever having to speak.
How can a deaf or hard of hearing employee use a voice tool?
Every spoken answer should also appear as text, and the tool should not assume everyone can wear an earpiece. For someone with hearing aids, the screen is often the real way in, and it must offer the same permissions and the same actions.
Do employers have to adapt digital tools for disabled employees?
In France, article L5213-6 of the Labour Code requires employers to take appropriate measures for recognised disabled workers, unless the cost is disproportionate, and to make sure the software they need for their work is accessible. Refusing can amount to discrimination. Other countries have comparable rules, such as reasonable adjustments in the UK.
Can a voice tool analyse how an employee speaks?
It never should. Scoring someone's diction, accent or hesitations has nothing to do with the job. And in the European Union, inferring emotions in the workplace has been banned since 2 February 2025, except for medical or safety reasons.