What 290 people with chronic illness told us about AI voices
To choose a voice for Juno, we tested 22 voices over six rounds with US adults living with long-term conditions. Here is every voice, how each one did, and what we learned.
Listen first
These are the three voices from our final check. Each one says the same three lines from a symptom check-in:
- “I've noted that you're feeling more tired today. When did that start?”
- “Was the headache mild, moderate, or severe?”
- “I've recorded the headache and tiredness for today. Would you like to add anything else?”
Play each voice, then pick the one you'd want. We'll show you which is which, and what 171 people chose.
58 of 171 picked it first · 5.6 out of 7
73 of 171 picked it first · 5.7 out of 7
40 of 171 picked it first · 5.5 out of 7
What we found
No voice clearly beat today's.
In the final check, the best new voice scored 5.7 out of 7 and today's voice 5.6. The gap was 0.1, with a likely range of −0.2 to +0.3. None of the three reached our bar of nine in ten people rating it 5 or higher: 84%, 83% and 81%.
A single rating is mostly noise.
Everyone in the final check heard their first voice again at the end. 76 of 171 gave it exactly the same score, and 139 were within one point. Even so, scores for the same voice varied by 1.2 points, almost as much as the 1.4 they varied between different voices. Only about a quarter of the spread looks like real taste.
Today's voice sounded cold more than slow.
People in our focus group had said our voice was slightly too slow. In the final check, only 26 of 171 people called it too slow, but 69 said it had too little warmth, against 46 for the best new voice. That gap is statistically clear (p < 0.01).
Asking for warmth, or for speed, backfired.
Voice G was made from the prompt “a warm, caring voice with a gentle smile in it, like a kind nurse checking in”. It came second to last in its round, and 14 of 24 people said it had too little warmth. One wrote: “She sounds so depressed.” Telling a stock voice to be warmer raised “too little warmth” from 6 to 11 of 24. Telling a voice to speak quicker made it faster, from 185 to 203 words a minute, and more people called it too slow: 15 of 24, against 10 for the same voice without the instruction.
Robotic and slow were the top complaints.
When people told us why they liked a voice least, the two most common reasons were that it sounded robotic or fake (56 of 271 reasons) or too slow (52).
Tired doesn't mean slower.
Of the 275 people who chose a voice both times, 232 (84%) picked the same voice for a day when they're exhausted as for an ordinary day. The 43 who switched moved to a quicker voice about as often as a slower one, 22 to 20.
Words in a voice prompt don't always do what they say.
Across 23 trial voices, the ones we described as low-pitched averaged 186 Hz, and the ones described as high-pitched averaged 184 Hz. In round 1, the voice we asked to be high came out low, and the one we asked to be low came out high. And in round 5, five voices made from exactly the same prompt scored anywhere from 4.4 to 5.8. One of them became our best new voice, and another was the least liked in its round.
Every voice we tested
All 22 voices, with the exact prompt behind each one. Eighteen were made with GPT-Live's prompt-to-voice, and two are OpenAI's stock voices, marin and gleam. Voices P and Q are earlier voices with one line added to their call instructions.
- Today's voice
OpenAI's stock GPT-Live voice
marin. No voice prompt. - A Low, calm
A low pitch for a feminine voice, in the alto range. Gentle, quiet energy with a smooth, soft texture; unhurried but never slow. Feminine presentation. Speak with a neutral General American English accent. An adult in their 30s. Natural everyday conversational pace, clear articulation, moderate vocal effort, moderately distinctive.
- B Low, plain
We asked for a high voice. It came out low.
A high-pitched feminine voice, speaking around 215 Hz. Direct and even, like a capable colleague; warm but never sweet or sing-song, with a clear, neutral timbre. Feminine presentation. Speak with a neutral General American English accent. An adult in their 30s. Natural everyday conversational pace, clear articulation, moderate vocal effort, moderately distinctive.
- C Low, bright
A low-pitched feminine voice in the alto range, speaking around 155 Hz. Light, upbeat energy with a clear, open timbre; friendly, not bubbly. Feminine presentation. Speak with a neutral General American English accent. An adult in their 30s. Natural everyday conversational pace, clear articulation, moderate vocal effort, moderately distinctive.
- D High, calm
A high pitch for a feminine voice, clearly above average. Gentle, quiet energy with a smooth, soft texture; unhurried but never slow. Feminine presentation. Speak with a neutral General American English accent. An adult in their 30s. Natural everyday conversational pace, clear articulation, moderate vocal effort, moderately distinctive.
- E High, plain
We asked for a low voice. It came out high.
A low-pitched feminine voice in the alto range, speaking around 155 Hz. Direct and even, like a capable colleague; warm but never sweet or sing-song, with a clear, neutral timbre. Feminine presentation. Speak with a neutral General American English accent. An adult in their 30s. Natural everyday conversational pace, clear articulation, moderate vocal effort, moderately distinctive.
- F High, bright
A high pitch for a feminine voice, clearly above average. Light, upbeat energy with a clear, open timbre; friendly, not bubbly. Feminine presentation. Speak with a neutral General American English accent. An adult in their 30s. Natural everyday conversational pace, clear articulation, moderate vocal effort, moderately distinctive.
- G Warm
A warm, caring voice with a gentle smile in it, like a kind nurse checking in; sincere and attentive, never sweet or sing-song. Feminine presentation. Speak with a neutral General American English accent. An adult in their 30s. Natural everyday conversational pace, neither slow nor rushed, clear crisp articulation, even volume, moderately distinctive.
- H Friendly
A friendly, lightly upbeat voice with natural rhythm, like someone glad to help; relaxed and human, never bubbly. Feminine presentation. Speak with a neutral General American English accent. An adult in their 30s. Natural everyday conversational pace, neither slow nor rushed, clear crisp articulation, even volume, moderately distinctive.
- I Steady
A calm, confident, professional voice with quiet warmth; clear and reassuring, never cold or flat. Feminine presentation. Speak with a neutral General American English accent. An adult in their 30s. Natural everyday conversational pace, neither slow nor rushed, clear crisp articulation, even volume, moderately distinctive.
- J Conversational
A natural, conversational voice with small easy pauses between phrases and questions that sound like real questions; relaxed, warm, and down to earth. Feminine presentation. Speak with a neutral General American English accent. An adult in their 30s. Natural everyday conversational pace, neither slow nor rushed, clear crisp articulation, even volume, moderately distinctive.
- K Friendly, long prompt
An original fictional adult voice, perceived age 30 to 40, with feminine presentation. Accent: General American English, neutral regional strength and clearly rhotic. Pitch: middle. Timbre: warm and clear. Texture: smooth, with a natural, unpolished conversational quality. Use moderate, easy conversational projection and a steady medium pace. Individual character: friendly and lightly upbeat, like someone glad to help, with natural rises on questions. Sound like a real adult speaking naturally to one person. Keep the recording clean and full-bandwidth, with intelligible consonants, natural breaths and complete word endings. Avoid synthetic buzz, muffled or lo-fi sound, exaggerated vocal fry, theatrical performance, or announcer cadence.
- L Bright, long prompt
An original fictional adult voice, perceived age 30 to 40, with feminine presentation. Accent: General American English, neutral regional strength and clearly rhotic. Pitch: middle-high. Timbre: bright and open. Texture: smooth, with a natural, unpolished conversational quality. Use moderate, easy conversational projection and a steady medium pace. Individual character: awake and engaged, with flexible, lively conversational intonation. Sound like a real adult speaking naturally to one person. Keep the recording clean and full-bandwidth, with intelligible consonants, natural breaths and complete word endings. Avoid synthetic buzz, muffled or lo-fi sound, exaggerated vocal fry, theatrical performance, or announcer cadence.
- M Grounded, long prompt
An original fictional adult voice, perceived age 30 to 40, with feminine presentation. Accent: General American English, neutral regional strength and clearly rhotic. Pitch: middle. Timbre: full and warmly resonant. Texture: smooth, with a natural, unpolished conversational quality. Use moderate, easy conversational projection and a steady medium pace. Individual character: kind and down to earth, attentive and capable, with small natural inflections. Sound like a real adult speaking naturally to one person. Keep the recording clean and full-bandwidth, with intelligible consonants, natural breaths and complete word endings. Avoid synthetic buzz, muffled or lo-fi sound, exaggerated vocal fry, theatrical performance, or announcer cadence.
- N Gleam
OpenAI's stock GPT-Live voice
gleam. No voice prompt. - P Friendly, told to be quicker
Voice H, with this added to its call instructions:
Keep a quicker everyday pace with no slow stretches, as if chatting with a friend.
- Q Gleam, told to be warmer
Voice N (
gleam), with this added to its call instructions:Sound calm, relaxed and genuinely at ease with this person, with more warmth and a little more variety in your tone.
- R Friendly prompt, new draw
The same prompt as voice H, made again.
A friendly, lightly upbeat voice with natural rhythm, like someone glad to help; relaxed and human, never bubbly. Feminine presentation. Speak with a neutral General American English accent. An adult in their 30s. Natural everyday conversational pace, neither slow nor rushed, clear crisp articulation, even volume, moderately distinctive.
- S Friendly prompt, new draw
The same prompt as voice H, made again.
A friendly, lightly upbeat voice with natural rhythm, like someone glad to help; relaxed and human, never bubbly. Feminine presentation. Speak with a neutral General American English accent. An adult in their 30s. Natural everyday conversational pace, neither slow nor rushed, clear crisp articulation, even volume, moderately distinctive.
- U Friendly prompt, new draw
The same prompt as voice H, made again.
A friendly, lightly upbeat voice with natural rhythm, like someone glad to help; relaxed and human, never bubbly. Feminine presentation. Speak with a neutral General American English accent. An adult in their 30s. Natural everyday conversational pace, neither slow nor rushed, clear crisp articulation, even volume, moderately distinctive.
- V Friendly prompt, new draw
The same prompt as voice H, made again.
A friendly, lightly upbeat voice with natural rhythm, like someone glad to help; relaxed and human, never bubbly. Feminine presentation. Speak with a neutral General American English accent. An adult in their 30s. Natural everyday conversational pace, neither slow nor rushed, clear crisp articulation, even volume, moderately distinctive.
- W Friendly prompt, new draw
The same prompt as voice H, made again.
A friendly, lightly upbeat voice with natural rhythm, like someone glad to help; relaxed and human, never bubbly. Feminine presentation. Speak with a neutral General American English accent. An adult in their 30s. Natural everyday conversational pace, neither slow nor rushed, clear crisp articulation, even volume, moderately distinctive.
Results by round
Score is the average answer to “How suitable is this voice for an assistant you would use to record your symptoms?”, from 1 to 7. The other columns count people.
Round 1, 25 people
| Voice | Score | Rated 5+ | Picked first | Liked least | Too slow | Too little warmth |
|---|---|---|---|---|---|---|
| Today's voice | 5.5 | 84% | 5 | 1 | 0 | 7 |
| E High, plain | 5.2 | 72% | 5 | 4 | 7 | 12 |
| D High, calm | 5.2 | 76% | 3 | 5 | 5 | 8 |
| C Low, bright | 5.2 | 68% | 2 | 4 | 1 | 10 |
| A Low, calm | 5.0 | 76% | 5 | 3 | 5 | 8 |
| B Low, plain | 4.9 | 60% | 1 | 4 | 8 | 17 |
| F High, bright | 4.6 | 52% | 4 | 3 | 15 | 10 |
Round 2, 24 people
| Voice | Score | Rated 5+ | Picked first | Liked least | Too slow | Too little warmth |
|---|---|---|---|---|---|---|
| Today's voice | 5.6 | 79% | 9 | 1 | 1 | 6 |
| H Friendly | 5.5 | 83% | 6 | 0 | 4 | 6 |
| J Conversational | 5.0 | 62% | 3 | 4 | 6 | 13 |
| E High, plain | 4.8 | 62% | 0 | 0 | 7 | 13 |
| D High, calm | 4.7 | 62% | 4 | 0 | 3 | 12 |
| G Warm | 4.4 | 46% | 2 | 7 | 7 | 14 |
| I Steady | 3.8 | 33% | 0 | 12 | 14 | 17 |
Round 3, 23 people
| Voice | Score | Rated 5+ | Picked first | Liked least | Too slow | Too little warmth |
|---|---|---|---|---|---|---|
| H Friendly | 5.7 | 83% | 8 | 0 | 6 | 2 |
| Today's voice | 5.7 | 78% | 6 | 4 | 0 | 5 |
| N Gleam | 5.6 | 83% | 2 | 1 | 2 | 4 |
| J Conversational | 5.5 | 87% | 1 | 2 | 7 | 5 |
| K Friendly, long | 4.5 | 57% | 0 | 3 | 4 | 11 |
| M Grounded, long | 4.2 | 35% | 4 | 2 | 2 | 10 |
| L Bright, long | 3.8 | 43% | 1 | 10 | 8 | 17 |
Round 4, 24 people
| Voice | Score | Rated 5+ | Picked first | Liked least | Too slow | Too little warmth |
|---|---|---|---|---|---|---|
| S New draw | 6.0 | 92% | 7 | 1 | 2 | 3 |
| N Gleam | 5.8 | 83% | 5 | 3 | 4 | 6 |
| H Friendly | 5.8 | 88% | 1 | 2 | 10 | 3 |
| Today's voice | 5.8 | 92% | 4 | 4 | 2 | 7 |
| R New draw | 5.5 | 79% | 3 | 1 | 3 | 12 |
| Q Told warmer | 5.5 | 71% | 2 | 6 | 7 | 11 |
| P Told quicker | 5.2 | 75% | 1 | 2 | 15 | 11 |
Round 5, 23 people
| Voice | Score | Rated 5+ | Picked first | Liked least | Too slow | Too little warmth |
|---|---|---|---|---|---|---|
| S New draw | 5.8 | 91% | 5 | 0 | 2 | 6 |
| V New draw | 5.8 | 96% | 4 | 1 | 6 | 5 |
| Today's voice | 5.7 | 87% | 2 | 0 | 4 | 5 |
| N Gleam | 5.7 | 87% | 4 | 1 | 3 | 6 |
| H Friendly | 5.6 | 83% | 6 | 1 | 10 | 6 |
| W New draw | 5.2 | 74% | 1 | 3 | 3 | 13 |
| U New draw | 4.4 | 48% | 0 | 10 | 2 | 16 |
Final check, 171 people
| Voice | Score | Rated 5+ | Picked first | Liked least | Too slow | Too little warmth |
|---|---|---|---|---|---|---|
| S New draw | 5.7 | 84% | 73 | 34 | 19 | 46 |
| Today's voice | 5.6 | 83% | 58 | 53 | 26 | 69 |
| V New draw | 5.5 | 81% | 40 | 84 | 59 | 53 |
How the study worked
Who took part. US adults living with a long-term health condition, recruited on Prolific. Everyone was paid above Prolific's recommended hourly rate, between £11 and £17 an hour at the median time. People were told that short quotes from their written answers might be published. We counted answers from 290 people.
What they heard. Each voice said the same three lines from a symptom check-in, recorded through GPT-Live with the same call instructions (voices P and Q each had one line added). Voices played in a balanced order under neutral labels, so no voice was always first, and nobody knew which voice Juno uses today.
What they answered. For each voice, people rated how suitable it was for an assistant they'd use to record their symptoms, how easy it was to follow, and how respectful and how natural it sounded, each from 1 to 7. They said whether its speed was too slow or too fast, and whether it had too little or too much warmth. Then they picked a favourite, a least favourite, and a voice for a day when they're exhausted.
How it ran. Five rounds of about 24 people tested seven voices each. After each round we kept the best voices and rewrote the rest. Then a final check with 171 new people compared today's voice with the two best new ones. Everyone had to choose, and the first voice they heard played again at the end so we could measure how much a single rating moves.
How we compared voices. We compared each voice with today's using each person's own scores for both. In the final check, the likely ranges are wider because we made three comparisons at once. Head to head, the best new voice won 87 of 154 choices against today's voice (56%, likely 47% to 66%), which is not a clear win.
What this can't tell us. These were short clips, heard once, by people on an online research panel in the US. Every voice was feminine, with a General American accent. A real call is longer, and the voice is only part of it. That's why the next step is a test on real calls.
What we're doing next
We're keeping today's voice for now. Next, we'll test the best new voice against it on real calls with people calling Juno for the first time, and see which one brings more people back.
We also want Juno to adapt to how you're feeling. Most people want the same voice on a hard day, so we'll keep the voice and change how Juno speaks instead, starting with fewer words rather than a slower pace. This was our first study of voice, and we'll be running more.
Talk to Juno yourself.
DownloadJuno: Chronic Illness Support is an app by SharedGenes, Inc. and is not affiliated with Juno Health (EHR software), Juno.com (email and internet service) or the film Juno.