README.md

metadata

tags:
  - qai-hub-models
  - full-duplex
  - speech-to-speech
  - qnn
  - snapdragon
license: other

ELAINAy/duplex-voice-models

Quantized full-duplex speech models compiled to Qualcomm QNN HTP context binaries for the Duplex Android app. The app reads catalog.json and downloads from the main branch, either from huggingface.co or a mirror such as hf-mirror.com.

Context binaries are tied to the listed SoC and QAIRT version and will not load elsewhere.

MiniCPM-o 4.5

9B 全模态模型的全双工语音部分（中英文）：Whisper-medium 听、Qwen3-8B 每秒决定听或说、TTS + CosyVoice2 流式发声。

Converted from openbmb/MiniCPM-o-4_5 by OpenBMB, license Apache-2.0.

Variant	Path	SoC	QAIRT	Size
LLM LPBQ W4 · Whisper / DiT / HiFT W8 · TTS CPU int8	minicpmo/minicpmo45-v1-sm8850	SM8850	2.45.0.260326	6.37 GB

Moshi · Moshiko

7B 全双工语音对话模型（英文，实验性），Mimi 24 kHz 编解码，80 ms 一帧，边听边说。

Converted from kyutai/moshiko-pytorch-bf16 by Kyutai, license CC-BY-4.0.

Variant	Path	SoC	QAIRT	Size
Temporal LPBQ W4 · DepFormer W8 · Mimi W8/W4	moshi/moshiko-lm-v4-lpbq4-sm8850	SM8850	2.45.0.260326	4.71 GB