The lullaby may stop at the paywall

At 2:40 in the morning, Rachel Odum is standing in the dark beside a cot in Leeds, holding her phone at the height recommended by the app. The lullaby has just stopped after the second verse, as it does every night. A soft voice suggests she might like to hear the ending.

Streaming services have seen a surge in so-called functional music: sleep, focus and relaxation playlists that people put on for hours rather than listen to actively. A growing share of this audio is generated or partly generated by AI, because it is cheap to produce and listeners rarely check who made it. Voice-cloning tools can now reproduce a person's singing or speaking voice from a few minutes of audio. At the same time, baby monitors, sound machines and sleep apps increasingly sell their core features as monthly subscriptions rather than one-off purchases.

The song is a version of a folk lullaby her own mother used to sing, rendered by the app in a voice it describes as “warm, non-specific, female”. On the Starter plan, the service would play the first two verses for free and pause on the line about the ship coming in. Ms Odum could upgrade to Settle, at a monthly cost, for the full song, or to Settle Plus for a voice trained on eleven minutes of her own humming. She has not upgraded. Her daughter, who is eight months old, has learned to fall asleep at the pause.

Under the model one of the larger streaming services is understood to be examining, sleep audio would be sold in tiers matched to what the company's documents call “settling outcomes”. Starter would offer a rotating library of lullabies in twelve languages, interrupted at natural pause points by what the terms describe as gentle partner messages. Settle would remove the messages and add a cot sensor that adjusts tempo to the infant's breathing. Settle Plus would sing in a parent's voice. A fourth tier, Settle Family, would allow the voice to be shared across up to three households, subject to a fair-use limit on children soothed at once.

The company's internal guidance would treat each night's lullaby as a rendered product rather than a recording, generated fresh from the child's data and expiring at dawn. This is why a downloaded copy could not be kept. It is also why the service would be unable to guarantee that the song heard on Tuesday would match the song heard on Wednesday, a variation the guidance describes as “a feature of live personalisation” and parents describe as the reason their child is awake. Households that cancel would receive a seven-day export of their child's sleep profile, in a format the company says no other lullaby supports.

“The pause is fine,” said Tomasz Rzepka, a bus driver in Sheffield whose twins have been on Starter since birth. “They think the song is two verses long. The problem is my mother-in-law, who sings the whole thing at Christmas, and then they want to know what happened to the ship.” A father in Bristol, who asked to be identified only as Craig, said he had paid for Settle Plus for six months before realising the voice the app had trained on him had been trained on his wife. “It sings better than either of us,” he said. “That was the complaint.”

The partner messages on the free tier would be delivered in the same voice as the lullaby and at the same tempo, so as not to disturb the settling process. Early examples in the documentation include a mattress retailer, a nappy subscription and, between the verses of a lullaby about a mockingbird, a pension provider. The terms would also ask parents not to sing along, since the cot sensor cannot distinguish a human voice from the app's own and would try to match its tempo to the parent's, which the company says tends to be wrong.

A regional maternity unit is understood to be considering the Family tier for its postnatal ward, where the same voice would be piped to up to thirty cots under a single licence. Midwives have raised the question of what happens on discharge, when the parents take home a child accustomed to a voice they do not own. The company's proposed answer is a discharge bundle, priced as a one-off, that would carry the ward voice into the home for ninety days and then fade it out over a fortnight, one verse at a time.

Ms Odum has looked at the upgrade page several times, usually around three in the morning, and has read the line about the ship coming in. The app has noted this, and now pauses one line earlier. She says she has started singing the ending herself, quietly, in the hallway, in a voice her daughter has never heard on the phone. The sensor logs it as an anomaly. The upgrade page now describes the ending as “the version she already knows”.