Nith Text-to-Speech

Nith Text-to-Speech#

ABR’s Nith family of Text-to-Speech (TTS) models use State Space Model (SSM) layers and were trained and released by Applied Brain Research (ABR). These model perform TTS with low latency from text buffer input to speech synthesis. These models also support features such as prosody and emotion control through SSML along with custom vocabulary pronunciation. With their extremely small parameter count and the efficiency of their SSM architecture, these models are able to deliver natural sounding speech in real-time on a wide variety of edge hardware platforms, from MCUs to GPUs.

Key Features#

  • Low-latency — Text is processed in small buffers mean vocalizations start quickly

  • Small footprint, high quality — Our models run right at the edge and offer big-model vocalization quality

  • Custom vocabulary — Specific pronunciation of brand names, technical jargon, and domain- or application-specific words can be instantly added with no model retraining

  • Multi-language — Available in English, Spanish, and Chinese (Mandarin), with additional languages in development and can be made available upon request

  • Output — 16KHz mono-channel audio (WAV)

Premium Features#

  • Custom voices — New TTS voices can be created from a 6-second audio clip

Models#

Model
Language
Parameters
nith-5m-live.en
English
5M
ArchitectureCompound CNN + SSM
QuantizationMixed precision 8-bit and 32-bit
nith-5m-live.es
Spanish
5M
ArchitectureCNN
QuantizationMixed precision 8-bit and 32-bit
nith-5m-live.zh
Chinese (Mandarin)
5M
ArchitectureCNN
QuantizationMixed precision 8-bit and 32-bit

About the Name#

ABR names its model families after rivers. State space models process sequential data as a continuous flow — always moving forward, maintaining state efficiently over time. The Nith model family is named for the river near ABR’s hometown of Waterloo, Ontario, Canada.

Contents