Khala (now Shao)

Paid APISelf-hostedWeb

What Khala (now Shao) is

Shao, formerly called Khala, is an open-source music-generation system that turns text descriptions and lyrics into complete, full-length musical works. Natural-language prompts and lyrics can guide style, mood, vocals, and content, including vocal generation. Its pipeline creates coarse acoustic tokens, completes higher layers in a 64-layer residual vector-quantization hierarchy with a super-resolution model, and decodes the waveform. The project includes a web frontend, FastAPI backend dispatcher, single-GPU inference worker, model loading, and an end-to-end audio path. Docker setup instructions provide a local web interface. The full pipeline recommends an NVIDIA GPU with at least 24 GB of VRAM, plus Docker, NVIDIA Container Toolkit, and a CUDA-compatible driver. The current release is aimed mainly at researchers and developers familiar with GPU servers and Docker and CUDA workflows. The linked model weights are listed under CC BY-NC 4.0, and commercial use is not permitted. Output quality can vary with GPU architecture and software versions, and bit-identical output across GPU types is not guaranteed.

Who it suits

It suits researchers and developers comfortable with GPU servers, Docker, and CUDA workflows. It is intended for non-commercial generation of complete musical works from prompts and lyrics.

What it does well

Generates full-length musical works
Prompt control for style and mood
Includes vocals generation
Local web UI and Docker setup
Open-source project

What to know first

Commercial use is not permitted
Requires NVIDIA GPU with 24 GB VRAM
Quality varies across GPU environments
Mainly aimed at technical users

Verdict

Shao provides a complete music-generation pipeline with prompt controls, vocals, and local deployment components. Its hardware, software, and non-commercial licensing requirements make it a specialized tool for technical users.

Facts from Khala (now Shao)'s own pages

Audience: The current release is mainly intended for researchers and developers familiar with GPU servers and basic Docker and CUDA workflows. · github.com, 12 Oct 2026
Commercial use: No · nanhuantech.com, 22 Sept 2026
Compatibility limitation: Generation quality can vary with GPU architecture and software versions, and the project says bit-identical output across GPU types is not guaranteed. · github.com, 12 Oct 2026
Deployment: The README provides Docker setup instructions and a local web UI at 127.0.0.1:30869. · github.com, 12 Oct 2026
Former name: The project was formerly called Khala and has been officially renamed Shao. · github.com, 12 Oct 2026
Full-song generation: It is designed to generate full-length musical works rather than short clips or loop-style accompaniment. · github.com, 12 Oct 2026
Generation architecture: Its pipeline uses a 64-layer residual vector quantization acoustic-token hierarchy, followed by a super-resolution model and decoder. · github.com, 12 Oct 2026
Generation controls: Natural-language prompts and lyrics can control style, mood, vocals, and content. · github.com, 12 Oct 2026
Generation pipeline: Its pipeline generates coarse acoustic tokens, completes higher RVQ token layers with a super-resolution model, and decodes the waveform. · github.com, 12 Oct 2026
GPU requirement: The full inference pipeline recommends an NVIDIA GPU with at least 24 GB of VRAM, such as an RTX 4090 or higher-tier GPU. · github.com, 12 Oct 2026
GPU requirements: The full inference pipeline recommends an NVIDIA GPU with at least 24 GB of VRAM, such as an RTX 4090 or higher-tier GPU. · github.com, 12 Oct 2026
Included system: The project includes a frontend UI, FastAPI backend dispatcher, single-GPU inference worker, model loading, and an end-to-end audio-generation path. · github.com, 12 Oct 2026
Intended users: The current release is mainly intended for researchers and developers familiar with GPU servers. · github.com, 12 Oct 2026
Model license: The linked model weights are listed as CC BY-NC 4.0. · huggingface.co, 12 Oct 2026
Model weights license: The model weights are stated to be intended for release under CC BY-NC 4.0. · github.com, 12 Oct 2026
Name change: The project was formerly called Khala and has been officially renamed to Shao, with its model, codebase, and research direction unchanged. · github.com, 12 Oct 2026
Supported platforms: web · nanhuantech.com, 22 Sept 2026
Prompt controls: Natural-language prompts and lyrics can control style, mood, vocals, and content. · github.com, 12 Oct 2026
Purpose: Shao, formerly Khala, is an open-source system for generating complete music from text descriptions and lyric conditions. · github.com, 12 Oct 2026
Quality limitation: The project says inference quality is sensitive to GPU architecture, CUDA stack, container image, and Megatron/TransformerEngine versions, and bit-identical output across GPU types is not guaranteed. · github.com, 12 Oct 2026
Runtime dependencies: The documented runtime uses Docker, NVIDIA Container Toolkit, and a CUDA-compatible NVIDIA driver. · github.com, 12 Oct 2026
Support: The README invites users to join a WeChat group for discussion, usage questions, and future updates. · github.com, 12 Oct 2026
System components: The system includes a frontend UI, FastAPI backend dispatcher, and single-GPU inference worker. · github.com, 12 Oct 2026
Verified environments: The released pipeline is reported to produce normal results on an H800 with the reference image and to work on an RTX 4090 when the provided environment is followed. · github.com, 12 Oct 2026

Where Khala (now Shao) stands on Freedom251

AI Music Generators#121 of 123