More and more, teams building voice-first products in AR glasses, robotics, automotive, and consumer devices want both ends of the voice I/O to live on the device. The reasons are the ones we covered in our recent post on cloud voice economics: cost, latency, and the architectural problems of routing every customer interaction through the network. The complication has been that on-device TTS has historically meant a compromise on quality and on production-grade features. That is now changing.
The parameter counts behind production text-to-speech systems have been falling fast. In March 2026, ElevenLabs, a leader in cloud TTS, established the production reference for on-device TTS with a model rumored to have hundreds of millions of parameters. Supertonic V3 launched in May 2026 at 99 million parameters, a significant compression that drew attention from the on-device speech community. In June 2026, ABR released its Nith family of TTS models which now take performant TTS to a new level at 5.5 million parameters. That is roughly 18 times smaller than Supertonic V3.
There are three things worth evaluating when judging a model this small: how it sounds, how it performs, and what customizations and production controls ship around it. Audio samples are embedded throughout so readers can judge quality directly rather than through a benchmark number.
The Parameter Race Reaches the Edge
The market trajectory is clear. TTS parameter counts have been coming down because production deployment economics demand it. Voice products that run on edge silicon, mobile devices, AR glasses, and embedded systems cannot accommodate models in the hundreds of millions of parameters. ABR’s Nith family is at the leading edge of that trajectory. Here are a couple audio samples showing how Nith output compares to the same phrase output from Supertonic V3:
Audio clip A: ABR Nith English TTS sample output speech
Audio clip B: Supertonic V3 English sample output speech (same text input as Nith)
How is 5.5 million possible? At ABR we are obsessed with pushing the limits of small, capable AI for edge processors. Our design discipline started from the smallest viable model and grew the architecture up to meet our quality targets, rather than compressing a large model down. That direction matters: starting small and growing up produces a structurally different model than starting large and shrinking down.
The runtime memory footprint is as small as 6 MB when quantized. That fits comfortably in MCU and NPU SRAM, where a 99M-parameter model does not. And a model of hundreds of millions of parameters is completely out of the question.
Paired with on-device ASR (ABR’s Niagara 38M ASR), the Nith family permits the low-latency and cost savings we described in our recent post on cloud voice economics. Both ends of the voice I/O can now sit inside the device’s power envelope and provide a fast, responsive user experience.
Small Doesn’t Have to Mean Robotic
Small TTS models have historically sounded mechanical. Achieving a natural sounding voice is the hurdle a 5.5 million parameter model has to clear. Naturalness has to be proven to the human ear. The audio samples in the previous section are the test. We invite you to listen and judge for yourself. ABR offers a number of natural sounding voices that produce realistic sounding speech and are available in multiple languages. Here is a sample of ABR Nith Spanish TTS:
Audio clip C: ABR Nith Spanish TTS sample output speech
And to be natural, TTS also has to be really responsive. ABR Nith TTS produces a first audible voice token in as little as 147 milliseconds on edge silicon – and it does this much more reliably than any cloud solution. This is worth emphasizing: because Nith models run on-device, variable network conditions don’t affect its performance.
For sustained streaming output, the real-time factor (RTF) determines how quickly continuous speech is produced relative to playback. An RTF below 1.0 means the model produces audio faster than real-time, which is the requirement for streaming voice products that need to respond in a timely manner. On our benchmarks across various edge silicon CPUs and NPUs, the Nith TTS models consistently achieve RTF well below 1, even as low as 0.27, meaning they always stream in real-time.
Customizing the Voice of Your Product
Voice cloning has been available in cloud TTS services for several years. The new development is that on-device TTS now does it too, even in the 5.5 million parameter Nith TTS family. A target voice audio sample as small as six seconds, processed through ABR’s consent-based voice cloning pipeline, produces a cloned voice that sounds as natural as the Nith TTS stock voices and introduces no change in performance or size. To demonstrate this, here is a sample audio clip from an interview of Audrey Hepburn coupled with Nith English TTS using a custom voice trained from that sample (note that ABR’s voice cloning pipeline can only be used with properly consented audio samples – the samples below are for illustrative purposes only):
Audio clip D: Audio sample used for voice cloning
Audio clip E: ABR Nith English TTS sample output using custom voice derived from sample above
Voice cloning has evolved as a must have for device manufacturers who want to establish a voice for their brand. In addition, there is a strong desire to be able to control how a voice speaks, not just how it sounds. To that end, the Nith model family also supports run-time controls to deliver the synthesized speech to fit the needs of the application.
Standard SSML markup gives product teams direct control over prosody (including rate, pitch, and volume) directly in the text that is fed to the TTS model. In addition, inline emotion tags make it easy to control tone and delivery to end users.
Other TTS Must Haves
Text strings in normal operation are frequently going to contain more than just words in the text. For this purpose, text normalization renders dates, currencies, temperatures, units, and common abbreviations as natural speech rather than character-by-character readouts. This is required for any voice product that touches numeric or formatted content, such as retail, finance, scheduling, weather, or healthcare applications.
In addition, custom vocabulary handles per-deployment pronunciation overrides for product names, technical terms, and brand language. This allows creating domain-specific TTS behavior without the need to retrain the model.
And of course, users speak a variety of languages and want to hear their devices speak to them in their native tongues. Multiple languages are available today in the Nith TTS family, including English, Spanish, and Mandarin, with Japanese and Korean coming very soon. The list will grow fast to cover demand in Asia and Europe. Voice cloning carries across these languages (see the previous audio example), useful for branded voices that need to ship internationally without re-recording the source speaker in each language (although to have authentic local accents this is recommended).
The Path Forward
At 5.5 million parameters, the ABR Nith family sounds natural, fits in a compact memory footprint, delivers a first voice token in as little as 143 milliseconds on edge silicon, supports streaming generation with RTF as low as 0.27, handles SSML and text normalization for production content, supports cloned voices on six seconds of enrollment audio, and is available in growing number of languages. The on-device half of the voice loop no longer requires a compromise on quality, naturalness, or production-grade features in order to ship on edge silicon.
At ABR, we believe that natural-sounding voice synthesis at edge-deployable parameter counts is the foundation of the next generation of voice products. If you are building voice-first AR, robotics, wearables, automotive, or consumer voice products, try the demo. For evaluation running on your product hardware, the English Nith TTS model is available in an evaluation SDK that can be downloaded from the ABR developer portal at dev.appliedbrainresearch.com. Download it, evaluate it on your own hardware, and talk to us about the voice you want to bring to your product.
Where the Nith name comes from: ABR names its model families after rivers to reflect their real-time streaming nature. The TTS models discussed in this post are the Nith family, named after the Nith River that runs through ABR’s home watershed in Waterloo, Ontario. Our ASR family is named for the nearby Niagara river.