---
base_model: openbmb/VoxCPM2
license: apache-2.0
library_name: audio.cpp
tags:
- tts
- voxcpm2
- quantized
---

# VoxCPM2-GGUF

This is an AgentQuantix TTS bundle for
`openbmb/VoxCPM2` using `audiocpp_cli`. The bundle is the unit
of publication: download every required file listed below.

## Runtime

```bash
audiocpp_cli --task tts --family voxcpm2 --model voxcpm2-F16.gguf --backend best --text "Hello world" --out out.wav
```

Supported languages: ar, my, zh, zh dialects, da, nl, en, fi, fr, de, el, he, hi, id, it, ja, km, ko, lo, ms, no, pl, pt, ru, es, sw, sv, tl, th, tr, vi

Expected sample rate: runtime-reported; see quality.json

Speaker reference: optional

## Bundle files

| File | Role | Quant | Size |
| --- | --- | --- | ---: |
| `voxcpm2-F16.gguf` | primary | F16 | 4551.2 MiB |
| `voxcpm2-BF16.gguf` | primary | BF16 | 4551.2 MiB |
| `voxcpm2-Q8_0.gguf` | primary | Q8_0 | 2818.1 MiB |
| `voxcpm2-Q3_K.gguf` | primary | Q3_K | 1527.3 MiB |
| `voxcpm2-Q4_K.gguf` | primary | Q4_K | 1798.3 MiB |
| `voxcpm2-Q5_K.gguf` | primary | Q5_K | 2053.3 MiB |
| `voxcpm2-Q6_K.gguf` | primary | Q6_K | 2324.2 MiB |

## Quality results

| Check | Result |
| --- | --- |
| F16 round-trip WER/CER | 0.184127 |
| F16 real-time factor | 4.8391 |
| F16 automated gate | True |
| BF16 round-trip WER/CER | 0.161905 |
| BF16 real-time factor | 4.4 |
| BF16 automated gate | True |
| Q8_0 round-trip WER/CER | 0.161905 |
| Q8_0 real-time factor | 5.0 |
| Q8_0 automated gate | True |
| Q3_K round-trip WER/CER | 0.195238 |
| Q3_K real-time factor | 5.4 |
| Q3_K automated gate | True |
| Q4_K round-trip WER/CER | 0.111905 |
| Q4_K real-time factor | 5.0 |
| Q4_K automated gate | True |
| Q5_K round-trip WER/CER | 0.106349 |
| Q5_K real-time factor | 5.7 |
| Q5_K automated gate | True |
| Q6_K round-trip WER/CER | 0.161905 |
| Q6_K real-time factor | 5.0 |
| Q6_K automated gate | True |

## Voice cloning and consent

Do not clone or imitate a person's voice without explicit, lawful consent. Do
not use generated audio for impersonation, fraud, deception, harassment, or
privacy-invasive activity. Review the upstream model license and acceptable-use
terms before use.

## License

The upstream model license (apache-2.0) applies to
the model artifacts. audio.cpp retains its own license.