| license: apache-2.0 | |
| language: | |
| - en | |
| - zh | |
| base_model: | |
| - ASLP-lab/MeanVC2 | |
| pipeline_tag: voice-conversion | |
| tags: | |
| - voice-conversion | |
| - zero-shot | |
| - streaming | |
| - meanvc2 | |
| - gguf | |
| - crispasr | |
| library_name: ggml | |
| # MeanVC2 — GGUF (ggml-quantised) | |
| GGUF / ggml quantised weights of [`ASLP-lab/MeanVC2`](https://github.com/ASLP-lab/MeanVC2) for use with **[CrispASR](https://github.com/CrispStrobe/CrispASR)**. | |
| MeanVC2 is a robust low-latency streaming zero-shot voice conversion architecture developed by ASLP@NPU (Apache-2.0). It couples: | |
| 1. **Pretrained Streaming Fast-U2++ Conformer ASR** (160ms/80ms window) for speaker-invariant bottleneck feature extraction. | |
| 2. **Universal Timbre Token Encoder (UTTE)** featuring Global Timbre Memory (GTM) and Temporal Timbre cross-attention (TVT). | |
| 3. **1-NFE Flow Matching DiT** with Future-Receptive Chunking (FRC) and causal KV caching. | |
| 4. **Vocos Neural Vocoder** with ConvNeXt backbone and inverse STFT synthesis. | |
| ## Quantization Details | |
| - **`meanvc2-120ms-40ms-q4_k.gguf` (77 MB)**: Weight linear layers quantised to `Q4_K` / `Q6_K`; attention projections, conv kernels, and norms preserved in FP16/FP32. | |
| - **`meanvc2-120ms-40ms-q8_0.gguf` (111 MB)**: Weight linear layers quantised to `Q8_0`. | |
| - **`meanvc2-120ms-40ms-f16.gguf` (175 MB)**: Unquantised IEEE FP16 baseline (100% numerical parity). | |
| ## Quickstart with CrispASR | |
| ### CLI Voice Conversion | |
| ```bash | |
| crispasr-convert meanvc2-120ms-40ms-q4_k.gguf input_voice.wav target_voice.wav output.wav | |
| ``` | |