GGUF packages of Edge0/Audio8-TTS-Realistic-Air, pinned to 23c305cc013fde378edb459f538317043c993a74. Each file contains the 601M DualAR model, its complete 44.1 kHz audio codec, and the original configuration, tokenizer and processor files.
Use the audio.cpp runtime at commit f7ac446fb75e594a6913c846bea6c85b637f6ac0, with the audio8_tts family. Build it with CUDA for the tested GPU path. This is an audio.cpp-native GGUF package.
Five quantizations, with two additional filename aliases for Hub discovery. Q3_K_S and Q4_K_S are Hub-compatible filenames for the uniform native Q3_K and Q4_K packages. Each alias has exactly the same bytes and SHA256 as its original filename; download either filename for that quantization. The native tensor storage, eligibility rules and audio.cpp runtime are unchanged.
Native formats are Q4_K, Q6_K, Q3_K, Q2_K and Q8_0. The codec uses F16 in every quantized package. The 896-wide main model has 28 matrices eligible for K quantization, covering about 20.3% of main-model parameters; remaining unaligned matrices retain source precision, and lookup tables use F16. Q8_0 can quantize more of the main model. Compare the actual sizes below when choosing a file.
| File | Component | Quantization | Size | Status |
|---|---|---|---|---|
| Audio8-TTS-Realistic-Air-Q4_K.gguf | tts | Q4_K | 1,715,091,872 bytes | Available |
| Audio8-TTS-Realistic-Air-Q6_K.gguf | tts | Q6_K | 1,746,552,224 bytes | Available |
| Audio8-TTS-Realistic-Air-Q3_K.gguf | tts | Q3_K | 1,698,885,024 bytes | Available |
| Audio8-TTS-Realistic-Air-Q2_K.gguf | tts | Q2_K | 1,686,491,552 bytes | Available |
| Audio8-TTS-Realistic-Air-Q8_0.gguf | tts | Q8_0 | 1,495,693,728 bytes | Available |
| Audio8-TTS-Realistic-Air-Q3_K_S.gguf | tts filename alias | Q3_K | 1,698,885,024 bytes | Available |
| Audio8-TTS-Realistic-Air-Q4_K_S.gguf | tts filename alias | Q4_K | 1,715,091,872 bytes | Available |
audiocpp_cli --task tts --family audio8_tts \
--model Audio8-TTS-Realistic-Air-Q8_0.gguf \
--model-spec-override embedded --backend cuda --threads 4 \
--text "Welcome. It is wonderful to hear from you." --out output.wav \
--request-option max_tokens=512 --request-option seed=42 \
--request-option temperature=0.8 --request-option top_p=0.95 \
--request-option top_k=50
Published files passed full container/tensor checks, exact embedded-source checks, GPU loading, and two English speech smoke cases with local transcription checks. The original-weight reference package was compared with the source implementation using matching prompt tokens, prefill logits and codec inputs. These checks establish a limited release smoke test; no benchmark scores were measured for these GGUFs. Voice cloning, all 23 source languages and the full source context have not been validated here. Detailed reports and generated audio remain local.
Each variant is produced independently from the pinned source tensors. Codec conversion fuses weight-normalization pairs before storage conversion. runtime-info.json, model-spec.json and the generated SHA256 manifest record the package provenance.
Apache-2.0; preserve LICENSE and NOTICE. The original architecture is inspired by the DualAR design used in Fish Audio S2 Pro. This is a community conversion of the Edge0 checkpoint.