---
title: Best Sub-2B TTS for CPU-Only Kiosk Audio
seo_title: Best Sub-2B TTS for CPU-Only Kiosk Audio
description: Best Sub-2B TTS for CPU-Only Kiosk Audio. Compare published evidence
  with a practical review procedure for CPU-only kiosk speech and select a MOSS speech
  model suited to the task.
topic: Text-to-Speech Models
date: '2026-11-15'
lastmod: '2026-11-15'
author: Model Workflow Lab Editorial
author_type: Organization
faq:
- q: Is a 100M checkpoint always faster?
  a: Audio codec, runtime and hardware affect latency. Test the full application.
- q: Should short cues be stereo?
  a: That depends on playback hardware and the mix. Verify the final channel configuration
    with the actual speakers.
category: library
copyright: Copyright 2026 ModelWorkflowLab.com; subject to stated permissions and
  exceptions.
---
# Best Sub-2B TTS for CPU-Only Kiosk Audio

## Best TTS model under 2B parameters?

MOSS-TTS-Nano is the first MOSS checkpoint to test for a CPU-only kiosk. At about 100 million parameters, it supports CPU streaming and 48 kHz stereo output. MOSS-TTS-Realtime and MOSS-VoiceGenerator are 1.7B alternatives when the task calls for conversational streaming or a designed voice.

![Cold-start the kiosk, then Measure audible onset, then Check repeated playback](/assets/images/sub-2b-tts-for-cpu-kiosk-audio.svg)

## An accessible device with spoken instructions

The National Library Service documents digital audiobook players with built-in audio instructions and tactile controls. That is a concrete device precedent for voice guidance close to the user. A kiosk pilot can apply the same design principle while testing a different implementation and audio path. [Public reference](https://www.loc.gov/nls/services-and-resources/informational-publications/digital-audiobook-players/)

## What the evidence measures

Parameter count is a screening constraint. It does not describe decoder memory, startup delay or audio buffering. The MOSS Nano release documents CPU use. The M5-TTS authors report 2.91% Seed-EN WER and 1.37% Seed-ZH CER in their own evaluation, making it a relevant compact comparison if you can reproduce the test conditions. [Evaluation source](https://arxiv.org/abs/2611.10162)

## Compare the options

| Approach | Evidence to collect | Decision |
| --- | --- | --- |
| Nano, about 100M | Cold start, CPU load and audio continuity | Evaluate for local dynamic prompts |
| Realtime, 1.7B | Turn gaps and interrupted playback | Evaluate if users converse with the device |
| VoiceGenerator, 1.7B | Fit of a voice designed from a description | Evaluate during voice design, before deployment |

## Measure latency on the actual kiosk

Measure from button press to audible speech, with the production screen and audio stack running. Separate first playback after idle from playback during a busy session. For fixed prompts, include prerecorded files as a baseline. Keep synthesis for fields that change often enough to justify the runtime work.

## Frequently asked questions

### Is a 100M checkpoint always faster?

Audio codec, runtime and hardware affect latency. Test the full application.

### Should short cues be stereo?

That depends on playback hardware and the mix. Verify the final channel configuration with the actual speakers.

## Related reading

- [sub-2B TTS model comparison](/best-tts-model-under-2b-parameters/)
- [offline accessibility prompts](/sub-2b-tts-offline-accessibility-prompts/)

## Sources

- [MOSS-TTS-Nano official model page](https://www.moss.cn/en/models/moss-tts-nano)
- [Nano code and weights](https://github.com/OpenMOSS/MOSS-TTS-Nano)
- [Realtime model card](https://huggingface.co/OpenMOSS-Team/MOSS-TTS-Realtime)
- [VoiceGenerator model card](https://huggingface.co/OpenMOSS-Team/MOSS-VoiceGenerator)
- [M5-TTS research report](https://arxiv.org/abs/2611.10162)
- [NLS digital players](https://www.loc.gov/nls/services-and-resources/informational-publications/digital-audiobook-players/)