On-Device Text-to-Speech for Wearables
On-Device Text-to-Speech for Wearables
On-Device
Text-to-Speech for Wearables
Written by Sohaib Ahmad
Written by Sohaib Ahmad
Written by Sohaib Ahmad
Published on 18/11/2026
Published on 18/11/2026
Published on 18/11/2026


Glasses, earbuds, watches and phones, speaking without the cloud.
Glasses, earbuds, watches and phones, speaking without the cloud.
Glasses, earbuds, watches and phones, speaking without the cloud.
Smart glasses, rings and watches that have to ask the cloud how to say a sentence will always answer a beat late and sound robotic. On-device text to speech for wearables removes that trip.
The voice is generated on hardware the user carries, or the phone it’s connected to, so it keeps working in a tunnel, on a flight or on a hospital ward, and the words it speaks stay on that hardware.
NeuTTS-2M is our device-native speech model for this job. It has 62.4M active parameters (173.6M in total), runs on the CPU alone with no GPU or NPU, speaks seven languages, clones a voice from about three seconds of audio, and scored a 0.47% word error rate on 100 Seed-TTS English sentences in our 2026 internal benchmark.
This guide covers why device teams move speech off the cloud and the challenges it faces to be deployed on wearable hardware.
Smart glasses, rings and watches that have to ask the cloud how to say a sentence will always answer a beat late and sound robotic. On-device text to speech for wearables removes that trip.
The voice is generated on hardware the user carries, or the phone it’s connected to, so it keeps working in a tunnel, on a flight or on a hospital ward, and the words it speaks stay on that hardware.
NeuTTS-2M is our device-native speech model for this job. It has 62.4M active parameters (173.6M in total), runs on the CPU alone with no GPU or NPU, speaks seven languages, clones a voice from about three seconds of audio, and scored a 0.47% word error rate on 100 Seed-TTS English sentences in our 2026 internal benchmark.
This guide covers why device teams move speech off the cloud and the challenges it faces to be deployed on wearable hardware.
Why run text-to-speech on the device at all?
Why run text-to-speech on the device at all?
On-device text to speech generates audio on the product’s own processor instead of a cloud server. For a wearable, that takes the network out of every spoken reply, keeps the text on the device, keeps the voice working with no connection, and swaps a per-character bill for a fixed cost per unit.
Each of those four benefits matters more on a wearable than on a laptop or a call-centre server, because the wearable is small, mobile and personal.
On-device text to speech generates audio on the product’s own processor instead of a cloud server. For a wearable, that takes the network out of every spoken reply, keeps the text on the device, keeps the voice working with no connection, and swaps a per-character bill for a fixed cost per unit.
Each of those four benefits matters more on a wearable than on a laptop or a call-centre server, because the wearable is small, mobile and personal.
Fast voice latency
Fast voice latency
Wearables are designed to speak in short turns, a translated sentence, a navigation prompt, the answer to a question asked out loud. With cloud TTS, each turn waits for the request to leave the glasses, cross Bluetooth to the phone, travel over Wi-Fi or mobile data to a data centre, and come back as audio. Device teams we speak to report 1 to 2 seconds per turn in the field.
On the device, the wait is compute time alone. That matters because text to speech is the last stage of a voice pipeline that also runs speech recognition and a language model. Every millisecond TTS doesn’t spend on the network goes back to those stages.
Wearables are designed to speak in short turns, a translated sentence, a navigation prompt, the answer to a question asked out loud. With cloud TTS, each turn waits for the request to leave the glasses, cross Bluetooth to the phone, travel over Wi-Fi or mobile data to a data centre, and come back as audio. Device teams we speak to report 1 to 2 seconds per turn in the field.
On the device, the wait is compute time alone. That matters because text to speech is the last stage of a voice pipeline that also runs speech recognition and a language model. Every millisecond TTS doesn’t spend on the network goes back to those stages.
Text data stays on the device
Text data stays on the device
The text a wearable reads aloud is often the most personal data it handles, a private message, a medication reminder, a translated conversation with a doctor.
With cloud TTS, that text goes to the vendor’s servers, and a contract governs what happens to it there. In the EU that’s a data processing agreement under GDPR; in US healthcare it’s a HIPAA Business Associate Agreement.
On-device synthesis is structurally different, because the vendor never receives the text. Without outbound speech data flows, security and privacy reviews are significantly simplified. On-device TTS doesn’t make a product compliant by itself, since other data flows still count, but it removes speech from the list of things to defend.
The text a wearable reads aloud is often the most personal data it handles, a private message, a medication reminder, a translated conversation with a doctor.
With cloud TTS, that text goes to the vendor’s servers, and a contract governs what happens to it there. In the EU that’s a data processing agreement under GDPR; in US healthcare it’s a HIPAA Business Associate Agreement.
On-device synthesis is structurally different, because the vendor never receives the text. Without outbound speech data flows, security and privacy reviews are significantly simplified. On-device TTS doesn’t make a product compliant by itself, since other data flows still count, but it removes speech from the list of things to defend.
Works offline
Works offline
Wearables are used everywhere, underground trains, flights, cold stores, mountain trails, hospital wards with poor signal. When a cloud voice loses its connection, the product either goes silent or falls back to the operating system’s built-in voice.
On-device voice behaves the same with or without a signal. We demo NeuTTS running entirely on the CPU of a Samsung Galaxy S24 in airplane mode, with nothing going in or out of the phone.
Wearables are used everywhere, underground trains, flights, cold stores, mountain trails, hospital wards with poor signal. When a cloud voice loses its connection, the product either goes silent or falls back to the operating system’s built-in voice.
On-device voice behaves the same with or without a signal. We demo NeuTTS running entirely on the CPU of a Samsung Galaxy S24 in airplane mode, with nothing going in or out of the phone.
How cost can scale
How cost can scale
Most wearables are sold once. Cloud TTS charges every time the device speaks, so the voice bill grows with every active user while revenue per unit stays flat. Here’s a worked example at published rate:
Most wearables are sold once. Cloud TTS charges every time the device speaks, so the voice bill grows with every active user while revenue per unit stays flat. Here’s a worked example at published rate:
Input | Value |
|---|---|
| Fleet size | 100,000 pairs of glasses |
| Speech per device per day | 3,000 characters (about 3 minutes of audio) |
| Characters per year | 109.5 billion |
| Cloud cost at $33 to $43 per million characters | $3.6M to $4.7M per year, rising with usage |
| On-device cost | A licence per unit, paid once |
Double daily usage and the cloud line doubles. The on-device line doesn’t change.
Double daily usage and the cloud line doubles. The on-device line doesn’t change.


Yearly cloud TTS cost for 100,000 smart glasses at published rates.
Large voice TTS models can’t fit on most hardware
Large voice TTS models can’t fit on most hardware
Wearables are limited to three things cloud TTS providers take for granted: memory, sustained compute and battery. Rings, earbuds and basic fitness bands run microcontroller-class chips. Glasses with an application processor, smartwatches and the companion phone have far more headroom. Neural voices that sound human take hundreds of megabytes to gigabytes of RAM, which is why most voice that runs on the wearable itself either sounds dated or runs on the phone in the user’s pocket.
Every on-device TTS decision comes down to two questions, which processor in the product does the speaking, and how much of its memory and CPU time can speech have?
Wearables are limited to three things cloud TTS providers take for granted: memory, sustained compute and battery. Rings, earbuds and basic fitness bands run microcontroller-class chips. Glasses with an application processor, smartwatches and the companion phone have far more headroom. Neural voices that sound human take hundreds of megabytes to gigabytes of RAM, which is why most voice that runs on the wearable itself either sounds dated or runs on the phone in the user’s pocket.
Every on-device TTS decision comes down to two questions, which processor in the product does the speaking, and how much of its memory and CPU time can speech have?


Voice model size against total device RAM, on a log scale.
GPU is not the real constraint, CPU native synthesis is
GPU is not the real constraint, CPU native synthesis is
Most wearables and edge boards are not fitted with a GPU. Many have a neural processing unit (NPU), but each chip vendor ships its own, with its own toolchain and operator support. A voice model that depends on one vendor’s NPU ties the product to that chip family for its whole hardware generation.
The CPU is the one processor every device has. NeuTTS runs on the CPU alone, ships in GGML format, and has a C/C++ inference path. We run it on Raspberry Pi boards, NVIDIA Jetson modules, MacBooks and the Samsung Galaxy S24.
The constraint of running on the CPU is that speech shares it with everything else. Speed per thread therefore matters as much as accuracy.
Most wearables and edge boards are not fitted with a GPU. Many have a neural processing unit (NPU), but each chip vendor ships its own, with its own toolchain and operator support. A voice model that depends on one vendor’s NPU ties the product to that chip family for its whole hardware generation.
The CPU is the one processor every device has. NeuTTS runs on the CPU alone, ships in GGML format, and has a C/C++ inference path. We run it on Raspberry Pi boards, NVIDIA Jetson modules, MacBooks and the Samsung Galaxy S24.
The constraint of running on the CPU is that speech shares it with everything else. Speed per thread therefore matters as much as accuracy.
On-device TTS benchmark
On-device TTS benchmark
Model | Parameters | Languages | Size on disk | Peak RAM | RTF, 1 thread (lower is faster) | WER (Seed TTS) | Voice cloning on device | Runs on |
|---|---|---|---|---|---|---|---|---|
| NeuTTS 2M | 60M | 7 | 245MB | 610MB | 0.55 | <0.5% | Yes | CPU |
| Phonon Multilingual | 40M | 5 | 74.6MB | 311MB | 0.36 | <0.5% | No | CPU |
| Kokoro | 82M | Single | 354MB | 687MB | 0.6 | <0.5% | No | CPU |
| NeuTTS Max | 1.7Bn | 44 | API | API | <0.05 | <0.5% | API | GPU or Accelerator |
How we measured: We ran the internal benchmark on the first 100 Seed-TTS English test sentences, and scored a corpus wide word error rate using Whisper Large V3 Turbo. RTF is the mean compute time divided by audio duration on a single thread.
Built for wearables and smart glasses
Built for wearables and smart glasses
On-device text to speech for wearables runs in one of two places. Either the wearable synthesises speech itself, which needs an application processor and spare memory, or the companion phone synthesises it locally and streams the audio over Bluetooth.
Both keep the text off the cloud. NeuTTS-2M creates the voice on the phones and connects to the wearable through Bluetooth.
On-device text to speech for wearables runs in one of two places. Either the wearable synthesises speech itself, which needs an application processor and spare memory, or the companion phone synthesises it locally and streams the audio over Bluetooth.
Both keep the text off the cloud. NeuTTS-2M creates the voice on the phones and connects to the wearable through Bluetooth.
Smart glasses
Smart glasses
For Smart Glasses speech is the main output and tool for engagement. Users get live translation read into their ear, turn-by-turn directions, messages read aloud, and scene descriptions for people with low vision. All of these happen mid-conversation or mid-movement, so a one-second cloud delay is obvious.
A typical architecture keeps everything local. The glasses capture the microphone audio, the phone or computer runs recognition, the language model and NeuTTS, and the audio streams back to the glasses’ speakers. With seven languages in one model (English, Spanish, French, German, Korean, Chinese and Japanese), the same build handles offline translation in all of them, in a cloned brand voice.
For Smart Glasses speech is the main output and tool for engagement. Users get live translation read into their ear, turn-by-turn directions, messages read aloud, and scene descriptions for people with low vision. All of these happen mid-conversation or mid-movement, so a one-second cloud delay is obvious.
A typical architecture keeps everything local. The glasses capture the microphone audio, the phone or computer runs recognition, the language model and NeuTTS, and the audio streams back to the glasses’ speakers. With seven languages in one model (English, Spanish, French, German, Korean, Chinese and Japanese), the same build handles offline translation in all of them, in a cloned brand voice.
Hearables and earbuds
Hearables and earbuds
Earbuds have no room for a neural TTS model, so the phone does the voice synthesis. The main requirement for on device hearables is consistency. The spoken notifications, workout cues and assistant replies have to come out in the same brand voice.
Earbuds have no room for a neural TTS model, so the phone does the voice synthesis. The main requirement for on device hearables is consistency. The spoken notifications, workout cues and assistant replies have to come out in the same brand voice.
Wrist and fitness wearables
Wrist and fitness wearables
A runner on a mountain trail wants pace and heart-rate cues without checking a screen, and often without coverage. Those cues are full of numbers, times and units, so the TTS has to read “5:42/km” and “162 bpm” correctly with no hand-written rules. Text normalisation (how the model turns numbers, times and units into spoken words) matters as much as voice quality here.
A runner on a mountain trail wants pace and heart-rate cues without checking a screen, and often without coverage. Those cues are full of numbers, times and units, so the TTS has to read “5:42/km” and “162 bpm” correctly with no hand-written rules. Text normalisation (how the model turns numbers, times and units into spoken words) matters as much as voice quality here.
On-device vs. on-prem vs. cloud: which deployment fits?
On-device vs. on-prem vs. cloud: which deployment fits?
The right deployment depends on where the text is allowed to go, whether the product must work offline, and how cost should scale. On device suits products that must work without a signal and keep speech private. On prem suits regulated organisations that need server scale throughput inside their own walls. Cloud suits products with a reliable connection that want the largest model with no hardware to manage. One NeuTTS model family covers all three, with the same API and voices.
The right deployment depends on where the text is allowed to go, whether the product must work offline, and how cost should scale. On device suits products that must work without a signal and keep speech private. On prem suits regulated organisations that need server scale throughput inside their own walls. Cloud suits products with a reliable connection that want the largest model with no hardware to manage. One NeuTTS model family covers all three, with the same API and voices.
On-device | On-prem | Cloud API | |
|---|---|---|---|
| Where speech is generated | The product’s own CPU or companion phone | Your servers or private cloud | Neuphonic’s servers |
| Network needed | No | Your internal network only | Yes |
| Where the text goes | Stays on the device | Stays inside your infrastructure | Sent to Neuphonic, with zero data retention |
| Cost model | Licence per unit | Licence or platform fee | Usage-based |
| Hardware | CPU only, about 610MB peak RAM for NeuTTS-2M | CPU or consumer GPU; up to 50 concurrent streams on one RTX 4090 | None on your side |
| Neuphonic models | NeuTTS-2M (enterprise), NeuTTS Nano and NeuTTS Air (open weights) | NeuTTS family | NeuTTS-Max |
Many products use more than one. Glasses might speak on-device by default and hand long form reading to the cloud when a connection is available. Because the models share an API and voices, that’s one integration, not two. Our guide to on-prem vs cloud vs on-device TTS covers the trade-offs for each setup in detail.
Many products use more than one. Glasses might speak on-device by default and hand long form reading to the cloud when a connection is available. Because the models share an API and voices, that’s one integration, not two. Our guide to on-prem vs cloud vs on-device TTS covers the trade-offs for each setup in detail.
How to deploy NeuTTS on your device
How to deploy NeuTTS on your device
NeuTTS models ship as GGML/GGUF files that run on a CPU through a Python package or a C/C++ inference path, with no GPU and no network connection. The fastest way to evaluate is to run the open NeuTTS Nano or NeuTTS Air weights on your target hardware, measure peak RAM and real-time factor there, and then talk to us about NeuTTS-2M and porting to your chip.
NeuTTS models ship as GGML/GGUF files that run on a CPU through a Python package or a C/C++ inference path, with no GPU and no network connection. The fastest way to evaluate is to run the open NeuTTS Nano or NeuTTS Air weights on your target hardware, measure peak RAM and real-time factor there, and then talk to us about NeuTTS-2M and porting to your chip.
Supported platforms and integration languages
Supported platforms and integration languages
Formats: GGML/GGUF, with 4-bit and 8-bit quantised variants of the open models.
Languages: Python and a C/C++ inference path for embedded builds
Hardware we run on: Raspberry Pi, NVIDIA Jetson, Apple Silicon MacBooks and Samsung Galaxy S24, all on CPU.
Porting support: our engineers port, tune and integrate the models on your hardware, and benchmark them against your memory and latency targets.
Formats: GGML/GGUF, with 4-bit and 8-bit quantised variants of the open models.
Languages: Python and a C/C++ inference path for embedded builds
Hardware we run on: Raspberry Pi, NVIDIA Jetson, Apple Silicon MacBooks and Samsung Galaxy S24, all on CPU.
Porting support: our engineers port, tune and integrate the models on your hardware, and benchmark them against your memory and latency targets.
Open models you can inspect and adapt
Open models you can inspect and adapt
An open model lets engineers stress test the numbers themselves. NeuTTS Air is released under Apache 2.0, and NeuTTS Nano under the NeuTTS Open License 1.0. Both are on Hugging Face, with code on GitHub, where the repository has 6,300 stars as of November 2026. NeuTTS-2M and NeuTTS-Max are available under an enterprise licence. Every NeuTTS output carries an inaudible Perth watermark, so generated speech can be identified. Our guide to open-source speech models and licensing explains what each licence allows in a shipped product.
An open model lets engineers stress test the numbers themselves. NeuTTS Air is released under Apache 2.0, and NeuTTS Nano under the NeuTTS Open License 1.0. Both are on Hugging Face, with code on GitHub, where the repository has 6,300 stars as of November 2026. NeuTTS-2M and NeuTTS-Max are available under an enterprise licence. Every NeuTTS output carries an inaudible Perth watermark, so generated speech can be identified. Our guide to open-source speech models and licensing explains what each licence allows in a shipped product.
#ondeviceAI
#ondeviceAI
#ondeviceAI
#texttospeech
#texttospeech
#texttospeech
#wearables
#wearables
#wearables
#smartglasses
#smartglasses
#smartglasses
FAQ
Everything you need to know about on-device text-to-speech for wearables.
On-device text-to-speech turns text into spoken audio on the product’s own processor rather than on a cloud server. The text never leaves the device, speech keeps working without a connection, and there’s no network delay or per-character charge. NeuTTS-2M, for example, runs on the CPU alone in seven languages.
Yes, in one of two ways. Glasses with an application processor and spare memory can run a compact model themselves. More often, the companion phone runs the model locally and streams audio over Bluetooth. NeuTTS-2M needs about 610 MB of free RAM, so today it typically runs on the phone or a computer.
No. NeuTTS, Kokoro-82M and Phonon-31M all run on a CPU alone. Running on the CPU matters for wearables because most have no GPU, and the NPUs they do have differ between chip vendors. The trade-off is speed per thread: NeuTTS-2M takes 0.55 seconds of compute per second of audio on a single thread.
It depends on the engine type. In our 2026 benchmark, peak RAM was 311 MB for Phonon-31M, 687 MB for Kokoro-82M and 610 MB for NeuTTS-2M.
On-device TTS keeps the text being spoken on the device, so no speech data is sent to or stored by a vendor. That removes speech from your GDPR data transfers and processor agreements. It doesn’t make a whole product compliant by itself; any other data your product collects or sends still needs its own review.
Yes. Once the model is on the device, it needs no connection at all. We demo NeuTTS running on a Samsung Galaxy S24 in airplane mode. That makes on-device TTS the right choice for flights, underground transit, cold stores, field equipment and anywhere else the signal drops.
FAQ
Everything you need to know about on-device text-to-speech for wearables.
On-device text-to-speech turns text into spoken audio on the product’s own processor rather than on a cloud server. The text never leaves the device, speech keeps working without a connection, and there’s no network delay or per-character charge. NeuTTS-2M, for example, runs on the CPU alone in seven languages.
Yes, in one of two ways. Glasses with an application processor and spare memory can run a compact model themselves. More often, the companion phone runs the model locally and streams audio over Bluetooth. NeuTTS-2M needs about 610 MB of free RAM, so today it typically runs on the phone or a computer.
No. NeuTTS, Kokoro-82M and Phonon-31M all run on a CPU alone. Running on the CPU matters for wearables because most have no GPU, and the NPUs they do have differ between chip vendors. The trade-off is speed per thread: NeuTTS-2M takes 0.55 seconds of compute per second of audio on a single thread.
It depends on the engine type. In our 2026 benchmark, peak RAM was 311 MB for Phonon-31M, 687 MB for Kokoro-82M and 610 MB for NeuTTS-2M.
On-device TTS keeps the text being spoken on the device, so no speech data is sent to or stored by a vendor. That removes speech from your GDPR data transfers and processor agreements. It doesn’t make a whole product compliant by itself; any other data your product collects or sends still needs its own review.
Yes. Once the model is on the device, it needs no connection at all. We demo NeuTTS running on a Samsung Galaxy S24 in airplane mode. That makes on-device TTS the right choice for flights, underground transit, cold stores, field equipment and anywhere else the signal drops.
FAQ
Everything you need to know about on-device text-to-speech for wearables.
On-device text-to-speech turns text into spoken audio on the product’s own processor rather than on a cloud server. The text never leaves the device, speech keeps working without a connection, and there’s no network delay or per-character charge. NeuTTS-2M, for example, runs on the CPU alone in seven languages.
Yes, in one of two ways. Glasses with an application processor and spare memory can run a compact model themselves. More often, the companion phone runs the model locally and streams audio over Bluetooth. NeuTTS-2M needs about 610 MB of free RAM, so today it typically runs on the phone or a computer.
No. NeuTTS, Kokoro-82M and Phonon-31M all run on a CPU alone. Running on the CPU matters for wearables because most have no GPU, and the NPUs they do have differ between chip vendors. The trade-off is speed per thread: NeuTTS-2M takes 0.55 seconds of compute per second of audio on a single thread.
It depends on the engine type. In our 2026 benchmark, peak RAM was 311 MB for Phonon-31M, 687 MB for Kokoro-82M and 610 MB for NeuTTS-2M.
On-device TTS keeps the text being spoken on the device, so no speech data is sent to or stored by a vendor. That removes speech from your GDPR data transfers and processor agreements. It doesn’t make a whole product compliant by itself; any other data your product collects or sends still needs its own review.
Yes. Once the model is on the device, it needs no connection at all. We demo NeuTTS running on a Samsung Galaxy S24 in airplane mode. That makes on-device TTS the right choice for flights, underground transit, cold stores, field equipment and anywhere else the signal drops.
Bring Neuphonic into your product.
Building something with voice AI? Talk to us about your use case,
deployment requirements, and how Neuphonic can fit into your stack.
Building something with voice AI? Talk to us about your use case,
deployment requirements, and how Neuphonic can fit into your stack.
Solutions
Resources
Neuphonic © 2026

Solutions
Resources
Neuphonic © 2026

Solutions
Resources
Neuphonic © 2026

