🌊 TinyWave: Compact & Expressive Speech Language Models
TinyWave is a family of efficient 2B-parameter speech language models distilled from the 7B SPIRIT-LM teacher. These models support speech-to-speech and interleaved speech–text generation, optimized for real-time use on commodity hardware.
Built through layer-aligned knowledge distillation, TinyWave models retain 93–97% of their teacher’s performance while using only ⅓ of the parameters — ideal for use in voice agents, assistive technologies, and edge devices.
📖 Read the paper: Efficient Interleaved Speech Modeling through Knowledge Distillation (arXiv:2506.23670)
🌐 Demo & samples: tinywave-landing
💻 Code: github.com/mohammadmahdinoori/TinyWave
🔧 Model Variants
| Model | Modality | Tokenizer | Description |
|---|---|---|---|
tinywave/speech-base-2b |
Speech → Speech | spiritlm_base |
Base phonetic-only speech generation |
tinywave/speech-expressive-2b |
Speech → Expressive Speech | spiritlm_expressive |
Includes pitch + style tokens |
tinywave/interleaved-expressive-2b |
Text ↔ Speech (interleaved) | spiritlm_expressive |
Multimodal expressive generation |
tinywave/expressive-spirit-lm-interleaved-librilight |
Teacher (7B, interleaved) | spiritlm_expressive |
LoRA-corrected SPIRIT-LM for distillation |