---
license: mit
tags:
  - text-to-speech
  - voxcpm
  - gguf
  - ghana-tts
base_model: ghananlpcommunity/ghana-tts-36k
---

# ghana-tts-36k — GGUF quantized variants

GGUF conversions of [`ghananlpcommunity/ghana-tts-36k`](https://huggingface.co/ghananlpcommunity/ghana-tts-36k)
for use with [bluryar/VoxCPM.cpp](https://github.com/bluryar/VoxCPM.cpp).

**The `.gguf` files are backend-agnostic** — the same file runs on both a CPU-only
and a CUDA build of the inference engine. Only the *binary* you compile differs
between CPU and GPU deployment, not the weight file.

## Files

| File | Quant | AudioVAE | Size | Notes |
|---|---|---|---|---|
| `ghana-tts-36k-f32.gguf` | F32 | original | largest | Reference/max-accuracy baseline |
| `ghana-tts-36k-f16.gguf` | F16 | mixed | | |
| `ghana-tts-36k-f16-audiovae-f16.gguf` | F16 | f16 | | |
| `ghana-tts-36k-q8_0.gguf` | Q8_0 | mixed | | **Recommended for CPU** |
| `ghana-tts-36k-q8_0-audiovae-f16.gguf` | Q8_0 | f16 | | **Recommended for GPU (CUDA)** |
| `ghana-tts-36k-q4_k.gguf` | Q4_K | mixed | | Smallest, more accuracy loss |
| `ghana-tts-36k-q4_k-audiovae-f16.gguf` | Q4_K | f16 | smallest of the practical options | Fastest CPU model-only RTF; good compact CUDA option too |

## Which one to use

Based on [bluryar/VoxCPM.cpp](https://github.com/bluryar/VoxCPM.cpp)'s own published
benchmarks on the same `"voxcpm"` architecture family (not run on this exact
checkpoint — validate on your own audio before committing to one in production):

**CPU inference:**
- Default pick: **`ghana-tts-36k-q8_0.gguf`** — best full-pipeline RTF at this model scale on CPU.
- Smaller/faster, slight accuracy trade-off: `ghana-tts-36k-q4_k-audiovae-f16.gguf`.

**GPU (CUDA) inference:**
- Default pick: **`ghana-tts-36k-q8_0-audiovae-f16.gguf`** — best full-pipeline RTF at this model scale on CUDA.
- Smallest CUDA-friendly option: `ghana-tts-36k-q4_k-audiovae-f16.gguf`.

## Inference

This repo includes two scripts that build and run [VoxCPM.cpp](https://github.com/bluryar/VoxCPM.cpp) against these weights:

- `infer_cpu.sh` — builds a CPU-only `voxcpm_tts` and runs it against `ghana-tts-36k-q8_0.gguf`.
- `infer_gpu.sh` — builds a CUDA-enabled `voxcpm_tts` (requires an NVIDIA GPU + CUDA toolkit) and runs it against `ghana-tts-36k-q8_0-audiovae-f16.gguf`.

Usage (either script):

```bash
bash infer_cpu.sh "Text to synthesize" prompt.wav "Exact transcript of prompt.wav" out.wav
bash infer_gpu.sh "Text to synthesize" prompt.wav "Exact transcript of prompt.wav" out.wav
```

Swap in a different `.gguf` from this repo by editing the `MODEL_PATH` variable
at the top of either script.