Sub-2B TTS for Offline Accessibility Prompts
Start with MOSS-TTS-Nano for a CPU-oriented offline voice prompt workflow. Its repository documents roughly 100M parameters, multilingual voice cloning, and 48 kHz stereo generation. MOSS-TTS-Realtime and MOSS-VoiceGenerator are distinct 1.7B checkpoints intended for streaming interaction and voice design respectively. Nano repository · Model family
Turn the CPU demonstration into an accessibility test
The Nano project documents local CPU inference and streaming. A proposed accessibility pilot can use a local app that reads navigation cues after its network connection is disabled. Test the first prompt after startup and the next few prompts after the model is warm. Evaluate intelligibility in the actual listening environment.
| Checkpoint | Task fit | Verify on your device |
|---|---|---|
| Nano, about 0.1B | Lightweight local prompts | Cold start, first audio, clarity |
| Realtime, 1.7B | Streaming voice interactions | End-to-end turn latency |
| VoiceGenerator, 1.7B | Design a voice from a description | Match to the approved voice brief |
The Realtime card reports roughly 180 ms time to first byte after warm-up in its published setup. Use separate timers for model startup, first playable audio, and full prompt completion. Parameter count alone leaves memory usage and listener experience unresolved. Realtime model card · VoiceGenerator model card
A good pilot includes short instructions that differ in one critical word. Test that listeners recover the intended action before choosing a deployment route. See the sub-2B model guide and identifier checks.
English SEO keywords
sub-2B TTS, offline accessibility, MOSS-TTS-Nano, CPU streaming.
Frequently asked questions
Which checkpoint is the lightweight CPU starting point?
Evaluate MOSS-TTS-Nano using the documented CPU route.
Can I apply Realtime’s 180 ms figure to Nano?
Each number belongs to its checkpoint and documented test setup.
What should an offline test include?
Run with network access disabled after provisioning the required model and runtime files.