Locally narrating an 80-character audiobook on an 8GB GPU with KeenLore's full pipeline

autonoma_2042 · reddit · 2026-11-20

A sci-fi author built KeenLore, locally hosted audiobook narration software capable of handling a complex novel with 80+ characters, singing, music and whispering. It runs in one container (CUDA 13, PyTorch, llama.cpp, audio.cpp) on a Ryzen 5 7600 + 8GB T1000 (50W) + 96GB RAM, with MoE experts in system RAM.

The pipeline has five stages: text analysis (Qwen3.6-35B-A3B for speaker attribution, voice descriptions, emotive direction and grammar — beating Muse Glimmer/GRMR); voice design and cloning (FireRedTTS3, OmniVoice, Qwen3-TTS 1.7B, VoxCPM2); narration (a Fish Audio Drama 3 fine-tune that "nothing else comes close" to); singing (DiffSinger with five voicebanks, plus SoulX-Singer, TCSinger 2); and QA (Whisper + wav2vec2 alignment + WavLM voiceprints + OpenLID driving unit tests for human-sounding output). Demo audiobook at 9m47s in the video; weights are non-commercial.

Original post →

More from Multimodal

Multimodal channel →