Harsha001/Audio8-TTS-Realistic-0.6b-ROCmFPX

Audio8-TTS-Realistic-0.6b ROCmFPX

GGUF files of Edge0/Audio8-TTS-Realistic-0.6b, a small multilingual text-to-speech model, for the ROCmFPX fork of llama.cpp. The runtime is native C++ (ggml): no Python, no PyTorch. It works on CPU, Vulkan (AMD Strix Halo and other GPUs) and CUDA.

Files

File	Size	What it is
Audio8-TTS-Realistic-0.6b-Q8_0.gguf	0.53 GB	text model, 8-bit (best quality)
Audio8-TTS-Realistic-0.6b-Q6_K_ROCMFPX.gguf	0.45 GB	text model, ROCmFPX 6-bit (recommended)
Audio8-TTS-Realistic-0.6b-Q4_0_ROCMFP4_STRIX.gguf	0.35 GB	text model, ROCmFPX 4-bit for Strix Halo
mmproj-Audio8-TTS-Realistic-0.6b-BF16.gguf	0.58 GB	code predictor + audio decoder (needed by every text model)

You need one text model file and the mmproj file. The ROCmFPX files carry small per-tensor scales; they need the ROCmFPX fork (commit 8a705e2f7 or newer).

Run

Build ROCmFPX (Vulkan example):

git clone https://github.com/HarshaVardhan3002/ROCmFPX && cd ROCmFPX
cmake -B build -DGGML_VULKAN=ON -DCMAKE_BUILD_TYPE=Release
cmake --build build --target llama-tts -j

Speak:

build/bin/llama-tts -m Audio8-TTS-Realistic-0.6b-Q6_K_ROCMFPX.gguf -mm mmproj-Audio8-TTS-Realistic-0.6b-BF16.gguf -ngl 99 \
    -p "Hello there! This is a quick test of the new voice." --output out.wav

• Write the text in the language you want spoken. There is no language switch.
• -s <seed> gives the same audio every time.
• One call makes at most about 47 seconds of audio (1024 frames).
• Voice cloning from a reference recording is not supported yet in this runtime. The model speaks in its own voice.

How close is it to the original?

Checked against the PyTorch reference with the same codes fed in (teacher forcing), an English test sentence, 113 frames. "Semantic picks" counts how often the top choice for the main speech token is the same as in PyTorch.

Text model (mmproj BF16)	text logits cos	semantic picks	code predictor picks
F32 (mmproj F32)	0.9999993	114 / 114	1010 / 1017
BF16	0.999997	114 / 114	1008 / 1017
Q8_0	0.99987	113 / 114	963 / 1017
Q6_K_ROCMFPX	0.99949	105 / 114	930 / 1017
Q4_0_ROCMFP4_STRIX	0.9952	93 / 114	720 / 1017

The audio decoder matches PyTorch at cos 0.9999998. Whisper transcribes the free-running output of every file in this repo word for word (English, two sentences, several seeds). Other languages were not tested yet.

On this small model the stock Q8_0 is more exact than our ROCmFPX 8-bit, so we ship stock Q8_0 for best quality. The ROCmFPX files are for smaller size and the AMD fast paths.

Speed was only measured on a busy CPU so far. GPU numbers from the community are welcome, please share yours.

How it works

• Text model (Qwen2-like, 24 layers, a plain llama GGUF): reads the text and picks the main speech token of each audio frame (21.5 frames per second of audio)
• Code predictor (4 small layers, in the mmproj): fills in the other 9 codes of the frame
• Audio decoder (arktts codec, in the mmproj): turns the 10 codes per frame into 44.1 kHz audio

The sampling settings (top-k 50, top-p 0.9, temperature 0.7, and the RAS rule against repeats) follow the reference.

Credits and licenses

• Model: Edge0/Audio8-TTS-Realistic-0.6b, Apache 2.0
• Runtime: ROCmFPX, built on llama.cpp (MIT)