Who has Angel's voice?
The persons we love, we care for, we cherish. A parent who sang us to sleep. A partner who laughed at the same joke for decades. A child who is only beginning to say our names. Their voice is how we know them in the dark, on the phone, in a crowded room.
An estimated 97 million people around the world may benefit from communication aids because they cannot rely on natural speech1. The causes differ: neurodegenerative disease such as ALS, stroke, cancer of the throat or mouth, injury, and more. For many, the voice goes before the mind does. The person is still there, with the same thoughts, humor and love, and can no longer be heard.
Today, Candle Fire AI takes on the effort of helping these vulnerable people be heard.
Status: draft, updated 13 November 2026. This is a PRFAQ: a press release written for the day we hope to launch (May 1, 2027), followed by the questions a skeptical reader would ask. The release describes a goal, not a shipped product. Training and automatic evaluation are finished, and the fine-tuned voice now runs offline on a laptop. It reached the level of real recordings on voice similarity; details are in the Internal FAQ. Large-scale testing with other speakers and listeners has not yet been conducted. Every number in this paper was measured or cited. Every aim is labeled as an aim.
Press Release
Angel Voice: keep the voice of someone you love
A Mandarin voice trained on a person's own recordings, owned by their family, running offline on their own devices, including eye-tracking communication devices.
May 1, 2027 — Losing the ability to speak is hard. Hearing a machine speak for you in a stranger's voice makes it harder. In one study of 166 people with ALS, 60 to 70 percent of those whose disease began in the speech muscles had lost functional speech about three years after their first symptoms2. Many people with ALS later communicate with their eyes, through an eye-tracking device that speaks what they select.
Good tools exist. Apple's Personal Voice now supports Mandarin3, but Mandarin results are not yet as good as many people hope, and the voice stays on Apple devices. ElevenLabs offers a trained voice clone, free for many people with permanent voice loss, and it already works in Tobii Dynavox communication software6. But that voice lives in ElevenLabs' cloud: it cannot be downloaded, and the device must be online with an active account to speak7.
Candle Fire AI today introduces Angel Voice. From about half an hour of a person's recordings, it fine-tunes a Mandarin text-to-speech model on that one person, with a single aim: to sound as close to them as possible. The trained model belongs to the family. It runs offline on their own computer, needs no account or subscription, and is built to plug into eye-tracking communication devices as a standard Windows voice.
The process is private and inexpensive. In our first run, 387 recordings (about 34 minutes) trained a voice for less than twenty dollars of computing time. Recordings are kept in private, encrypted storage during training and are never sent to a third-party speech service.
Speaking with the eyes is slow: every letter or word takes a deliberate glance. So Angel Voice will also suggest whole sentences. As a person begins to select words, a language model proposes likely sentences, which they can accept with one glance and hear in their own voice.
"When my husband could no longer speak, the device that read his words aloud sounded like a stranger. Hearing him say 'good morning' in his own voice again would mean everything to our family," said a caregiver of a person with ALS.
Angel Voice is designed to reach people in three ways, from simplest to most flexible:
- A phrase library. A few hundred everyday phrases, such as greetings, requests and personal expressions, recorded in advance in the person's voice and loaded into any communication app that plays audio files.
- A home program. The voice runs on a family computer and turns any typed sentence into speech, with no internet connection.
- A voice for eye-tracking devices. Dedicated communication devices, such as Tobii Dynavox, run Windows and accept installed speech voices8. Angel Voice will install as one of them, so a person can select words with their eyes and hear them in their own voice, with no internet connection. This is a design, not yet built.
Angel Voice is a research project from Candle Fire AI, which also builds a biomedical knowledge graph for ALS research. It is not a medical device and makes no clinical claims.
Customer FAQ
Who is it for?
People who are losing, or have lost, the ability to speak, and the families who care for them. The first target is Mandarin speakers with ALS, because that is where we started and where the need to act early is greatest.
What do I need to do?
Record yourself, in a quiet room, reading varied sentences: statements, questions, exclamations, and some long sentences. About half an hour of clean speech is what our first run used. The sooner you record, the better, while your voice is at its strongest. We then handle cleaning, labelling and training.
Will it sound like me?
On our first speaker, the measurements are encouraging. On a speaker-similarity measure, the fine-tuned voice scored 0.962, against 0.955 for real recordings of the same person and 0.917 for the cloning method we started with. That measure is known to understate differences, though, and one speaker is not proof. The real test is human: people who know the voice, listening blind, across more speakers. We will publish the result whether it is good or bad.
Who hears my recordings?
You, and the computer that trains the voice, for the length of the run. Recordings and models sit in a private, encrypted cloud bucket with public access blocked, and are never sent to a speech service. A voice recording is biometric data, and we treat it as such.
How is it different from Apple, Microsoft and ElevenLabs?
All three already train a voice on a person's own recordings, and each is a good choice for some people. If you use Apple devices and the result sounds right to you, Personal Voice is the quickest way to start. If you speak English and your device is usually online, ElevenLabs' free program for people with voice loss is worth applying to today. The table compares them as of November 12, 2026, based on the vendors' own pages. These products change often, so check each vendor before deciding.
| Dimension | Apple Personal Voice | ElevenLabs Professional Voice Clone | Microsoft Azure professional voice | Angel Voice |
|---|---|---|---|---|
| Trained on your voice | Yes, overnight on the device4 | Yes, a dedicated model7 | Yes5 | Yes |
| What you record | A short guided session on an iPhone or iPad | At least 30 minutes; 2 to 3 hours recommended | At least 300 sentences (about 30 minutes) | About half an hour of clean, varied speech |
| Mandarin | Supported. No published Mandarin evaluation; quality is not yet as good as many expect | Listed as supported. No published Mandarin evaluation found | Supported (zh-CN). No published evaluation found | The first language, not an add-on. Close on similarity for one speaker; accuracy still needs work |
| Where the voice lives | Your Apple devices | ElevenLabs' cloud. Cannot be downloaded | Microsoft's cloud | Your own computer, owned by your family |
| Works offline | Yes, on Apple devices | No. Needs internet and an active account | No | Yes |
| Eye-tracking devices | Not on Windows devices | Yes, in Tobii Dynavox TD Snap and TD Talk, online with an account6 | Only through apps built on the Azure API | Planned, as an installed offline Windows voice. Not yet built |
| Who can use it | Anyone with a supported Apple device | Anyone with a paid plan; only for your own voice, after you verify it by speaking | Businesses, after Microsoft approves the use case5 | Research project; not yet available to the public |
| Cost to the family | Free | Subscription, or free through its Impact Program for people with permanent voice loss6 | Pay per use | No subscription. Training costs less than $20 |
We have not tested any of these products ourselves and have not yet run a side-by-side listening test. We plan to, and will publish the result.
Where Angel Voice stands apart: fine-tuning aimed at one goal, sounding as close to the person as possible in Mandarin; a voice the family owns and keeps, with no account to lapse; and offline speech on the eye-tracking devices many people with ALS end up depending on. Whether it also sounds closer than the others is exactly what the side-by-side test is for.
How do I use it in my communication device?
Start with the phrase library: no delay, works anywhere audio plays, and it covers most daily needs. For new sentences, the voice runs on a family computer. Today it is slow: on a recent laptop without using its graphics chip, 17.5 seconds of speech took 48 seconds to generate. Making it fast enough for conversation is part of the next stage. For eye-tracking and other dedicated Windows devices, such as Tobii Dynavox, Angel Voice is planned as an installed voice that works offline, together with sentence suggestions so a person needs fewer glances to say what they mean.
Internal FAQ
Why train a voice instead of cloning one from a short clip?
We tried cloning first, with Qwen3-TTS and IndexTTS-1.5. In these models the reference voice weakens as the speech gets longer: the start of a sentence sounds like the speaker and the end drifts toward a default voice. In our baseline, similarity on the long test prompts fell from 0.930 in the first third to 0.891 in the last third. Splitting text into 20-character pieces helps. Using each output as the next reference made it worse, because the errors compounded. Neither reaches the quality of a model that has learned the voice.
Apple and ElevenLabs already fine-tune. Why build our own?
Because closeness to the person is the whole point, and nobody has shown how close any of them gets in Mandarin. Apple's only published evaluation used 44 English speakers4; we found no published Mandarin evaluation for ElevenLabs or Microsoft. Owning the training lets us push further where it matters: more of the person's own recordings, careful transcript review, choosing checkpoints by how the voice sounds to family, and retraining as more recordings are found. It also lets the family keep the result, offline. If the side-by-side test shows another product already sounds closer, we will say so and point families to it.
Why CosyVoice 2?
It had the best Chinese quality of the options we compared, official fine-tuning scripts, and a design that avoids the same drift. GPT-SoVITS v3 was the close alternative; it needs more data to converge.
What have we actually measured?
| Stage | Result |
|---|---|
| Data | 387 training and 44 validation clips (34.3 and 3.5 minutes), merged from seven curated sets and checked by a human |
| Smoke test | Passed: one epoch of each training stage, 23 minutes, about $0.55 |
| Full training | Done: 6 epochs, 5,509 seconds, less than $20 |
| Evaluation | Done: 25 test sentences, four model variants, automatic scores (below) |
| Local voice | Running offline on a laptop (Apple M4, 16 GB, processor only). Short and medium test sentences: no character errors, similarity 0.98. Speed: 17.5 seconds of speech took 48 seconds |
What were the results?
Scores on 25 test sentences never used in training. Similarity is speaker similarity (higher is closer; real recordings of the speaker score 0.955). CER is character error rate (lower is better).
| Model | Similarity | CER |
|---|---|---|
| Qwen3-TTS cloning (starting point) | 0.917 | 0.9% |
| CosyVoice 2, not fine-tuned | 0.927 | 8.5% |
| Fine-tuned, version chosen for local use | 0.962 | 5.0% |
Fine-tuning lifted similarity to the level of real recordings. CER is measured by a speech recognizer, and the speaker has a regional accent: on the speaker's own real recordings, the same recognizer misreads 8.6% of characters9. A voice that keeps the accent scores a higher CER for that reason. The starting point's low CER partly reflects that it sounded less like the speaker.
What are the biggest risks?
- Noisy audio. The model learns room noise as part of the voice. Mitigation: denoising and manual clip review.
- Wrong transcripts. A 3% error rate over 400 clips is about 50 bad pairs. Mitigation: human review, aimed at names, digits and sentence-final particles.
- Narrow range. A voice trained on calm statements may sound flat on questions or exclamations. Mitigation: record varied sentences on purpose.
- Interrupted training. Spot capacity can be reclaimed. Mitigation: a checkpoint every epoch.
- Small evidence. One speaker and one run prove little about other voices. We will not generalize until others have tried it.
- Too slow on eye-tracking devices. On a laptop processor the voice already takes about three times as long as the speech it produces, and eye-tracking devices are usually less powerful. Mitigation: measure on a typical device; speak sentence by sentence as text is ready; use a compressed or smaller model; fall back to the phrase library.
- Sentence suggestions that speak for the person. A language model may propose what it thinks is likely rather than what the person means. Mitigation: suggestions are never spoken without the person's explicit choice, and typing word by word always remains available.
- Others are already good enough. If another product sounds as close in Mandarin, our effort should go to helping families use it. Mitigation: the side-by-side test comes before any large investment.
What are we not claiming?
- That this is a medical device, or has been validated with patients or clinicians.
- That the voice is indistinguishable from the real one. That is the goal, and it is untested.
- That the eye-tracking device voice exists. It is a design.
- That Angel Voice sounds closer than Apple, ElevenLabs or Microsoft. That is untested.
- That the voice is ready for conversation. It is still too slow, and its accuracy still needs work.
- That sentence suggestions exist or help. They are the next stage, not yet built or tested.
- That families can set it up alone. Today it takes technical help.
What happens next?
Training, evaluation and offline use on a laptop are done. The next stage has two parts:
- Sentence suggestions with a language model. As a person selects words, propose a few likely complete sentences in Mandarin, which they can accept with one glance. We will measure how many glances it saves and how often the suggestions match what the person meant.
- Connecting to eye trackers. Package the voice as an offline Windows voice that eye-tracking communication software can select, measure its speed on a typical device, and talk with eye-tracking device makers, including those serving Mandarin speakers.
Alongside that work:
- Improve accuracy, and make speech fast enough for conversation.
- Report the blind listening test.
- Run a side-by-side Mandarin test against Apple Personal Voice and an ElevenLabs Professional Voice Clone, with the same speaker and recordings.
- Find one or two other Mandarin speakers willing to try it, so the result is not a sample of one.
Sources
- American Speech-Language-Hearing Association, Augmentative and Alternative Communication, citing Beukelman & Light (2020): "approximately 5 million Americans and 97 million people in the world may benefit from AAC." ASHA notes that AAC prevalence is hard to estimate. This is a count of people who may benefit from communication aids, not a count of people who have lost speech.
- Massachusetts General Hospital Institute of Health Professions, New research establishes rate of ALS speech decline. Most people whose ALS began in the limbs kept functional speech over the same period.
- Apple, iOS and iPadOS 26 feature availability: Personal Voice is listed for English (United States), Mandarin Chinese (China mainland) and Spanish (Mexico). Our statement that Mandarin quality is weaker than many expect is our own assessment.
- Apple Machine Learning Research, Personal Voice: models are fine-tuned on the user's device from 150 recorded sentences; the published evaluation used 44 adult English speakers.
- Microsoft Learn, Train your professional voice model and Text to speech FAQ (at least 300 recordings, about 30 minutes; zh-CN supported); What is personal voice? (access restricted to eligible customers and approved use cases).
- ElevenLabs, Impact Program and announcement; Tobii Dynavox, How do I use my Eleven Labs voice in TD Snap? (requires an API key, an active subscription and tokens). Impact Program terms and eligibility should be confirmed with ElevenLabs.
- ElevenLabs, Professional Voice Cloning: a dedicated model trained on 30 minutes to 3 hours of audio; Mandarin listed; only your own voice, after verification; the clone cannot be downloaded or exported.
- Tobii Dynavox, AAC voice options: personal voices can be installed on its Windows devices as SAPI5 voice files, with offline options depending on the provider.
- Our measurement: Qwen3-ASR 1.7B, the recognizer used in our evaluation, on 4,437 characters of the speaker's real recordings, compared with human-checked transcripts. Common confusions include f heard as h and l heard as n.
Contact
Questions, corrections, or an interest in trying this: linkan@candlefireai.org.