A default female voice binds the assistant role to gender
Aliases: gendered assistant · female-default TTS · assistant-gender binding
What it is
Most voice assistants ship with a female voice and spend their time fetching, querying, booking, turning lights on. The social gender of the voice and the subordinate structure of the task are bound in a single default. People need not believe there is a woman on the other side to run the script “the one who helps is female” — that is the CASA prediction for a gender cue. The binding sits on the default: not whether the system claims to be a woman, but that, until someone changes settings, assistant work is done in a female voice. The UNESCO-style critique of voice assistants is aimed at this bundle, not at whether the timbre is pleasant.
Why it happens
CASA starts from one social cue being enough to run an interpersonal script. The gender of a synthetic voice is a strong cue — it reaches the ear before a name or a pronoun. The assistant’s task structure is a second cue: wait to be told, do not push back, define success as “I handled it for you.” Taken together, listeners predict politeness, patience, and reaction to interruption the way they predict those things of women in service roles. The same script in a male voice moves competence ratings on authority tasks and “appropriateness” ratings on service tasks — not because the model got smarter, because a gender stereotype was woken by the sound. Choosing female as the factory voice repeats that conjunction in every session where settings are never touched. Advertising and the first unboxing utterance perform the default in public, and the binding walks from the device in the kitchen into the culture of what an assistant sounds like.
Wording, sentence length, and forms of address can be designed separately; they change style inside the same gender cue. Until the default gender is taken apart, a more “neutral” style is still a female voice waiting to be told.
Studying it
CASA gender experiments: the same task replies, only the TTS gender cue swapped; measure warmth, competence, fitness to give orders / fitness to take them. Split tasks into service (book, look up, lights) and authority (advise, refuse, explain a rule) and read the cue × task interaction, not “do you like this voice.”
On the product side, inventory defaults: factory voice, voice in ads, whether authority skills (bank adviser, medical instruction) get a male voice instead. Content analysis follows the UNESCO line: whether the bundle is repeated at factory and in promotion — without inventing market-share numbers. Lab preference cannot stand in for “who unaltered settings actually hear every day.”
Where it stops holding
A user who explicitly wants a female voice, or an accessibility user who depends on one familiar timbre, is choosing, not receiving a factory bundle. In languages without a stable vocal-gender opposition the cue weakens and binding moves onto address terms and occupational nouns. Child products in a child’s voice have age scripts overlaying gender scripts. Labeling every female voice a stereotype would drive women’s voices back out of professional and authority skills. The problem is the default locking female voice onto a subordinate role, not that a female voice cannot advise.
Applying it
- Do not ship “the one who waits to be told” as a female voice. Force a choice once, or rotate by day, until the user pins it.
- Do not divide authority and service skills by gender (adviser male, booking female). Name roles in the same product by duty, not by the social gender of the timbre.
- Ads and the first unboxing sentence should not use “female voice + waiting to serve” as the only demo, in line with the product default.
- How to check: an untouched new device speaks three service turns and three refusals. If all six are the same female voice waiting and apologizing, the binding is still on the factory value.