kizuna-intelligence

kizuna-intelligence / Irodori-TTS-Lite

Public

Int4-quantized inference runtime for Faster-Irodori-TTS2 voice-design DiT. ~1GB VRAM end-to-end.

46
2
85% credibility
Found May 20, 2026 at 46 stars -- GitGems finds repos before they trend. Get early access to the next one.
Sign Up Free
AI Analysis
Python
AI Summary

Irodori-TTS-Lite is a lightweight runtime that lets you run a Japanese text-to-speech model using only 1 GB of GPU memory instead of nearly 3 GB. It works by using special compressed model files that are 85% smaller on disk, while maintaining near-identical audio quality. You install it with one command, and it automatically hooks into your existing TTS system so you can type text and hear natural-sounding Japanese speech. The project includes tools to measure exactly how much memory your setup uses, and you can choose between faster performance or even lower memory usage depending on your hardware.

How It Works

1
๐ŸŽ™๏ธ You want to add voice to your project

You discovered that the original Japanese TTS model needs almost 3 GB of GPU memory, which is too much for your setup.

2
๐Ÿ“ฆ You install the lightweight version

You install Irodori-TTS-Lite, which automatically downloads the compressed model weights from the internet.

3
โœจ You connect it to your existing TTS system

With one simple command, your TTS system learns to read the compressed model files directly.

4
๐Ÿ’ฌ You type your text and hear the voice

You enter Japanese text like 'ใ“ใ‚“ใซใกใฏใ€ใƒกใƒฉใ ใ‚ˆใ€‚' and the system converts it to speech.

5
You choose your memory balance
๐Ÿš€
Speed mode

Keep everything on GPU for ~1 second generation time

๐Ÿ’พ
Memory saver

Move the codec to CPU, using only ~1 GB of GPU memory total

6
๐ŸŽต Your audio file is ready

The generated speech is saved as a WAV file, sounding nearly identical to the original uncompressed version.

๐ŸŽ‰ Everything works with a fraction of the resources

Your voice project now runs on just 1 GB of GPU memory instead of nearly 3 GB, with almost no loss in quality.

Sign up to see the full architecture

5 more

Sign Up Free

Star Growth

See how this repo grew from 46 to 46 stars Sign Up Free
Repurpose This Repo

Repurpose is a Pro feature

Generate ready-to-use prompts for X threads, LinkedIn posts, blog posts, YouTube scripts, and more -- with full repo context baked in.

Unlock Repurpose
AI-Generated Review

What is Irodori-TTS-Lite?

Irodori-TTS-Lite is a lightweight runtime for running a Japanese text-to-speech model with int4 quantization baked in. It takes the full Irodori-TTS2 pipeline and squeezes it down to under 1GB of VRAM for end-to-end inference, compared to nearly 3GB for the standard version. Built in Python, it uses custom Triton kernels to execute 4-bit quantized operations directly without decompressing the entire model first. You drop it into an existing Irodori-TTS setup with a single function call and it intercepts the checkpoint loading process automatically.

Why is it gaining traction?

The numbers are the story. Disk size drops from 1.88 GB to 279 MB. Peak GPU memory falls from 1.9 GB to 552 MB for the core model, or under 1 GB for the full pipeline including the codec. Audio quality stays nearly identical to the full-precision version. The killer feature is that this works as a transparent drop-in: call patch() once, and your existing inference code loads 4-bit checkpoints without any other changes. Optional flags let you push the codec to CPU for another 500 MB savings, or quantize the codec to int4 as well if latency tolerance allows.

Who should use this?

Developers running Japanese TTS on consumer GPUs with limited VRAM will get the most value. If you have an RTX 3060 or similar with 8GB, this frees up meaningful headroom for concurrent workloads. Research teams comparing quantization methods will appreciate the clear benchmarks and memory profiling tooling. Anyone already invested in the Irodori-TTS ecosystem can try it with zero rewrites, though you'll need CUDA and Triton support.

Verdict

This is a technically solid quantization solution for a specific niche, with a credibility score of 0.85% reflecting its early stage. At 46 stars it has minimal community validation, and documentation leans toward internal design notes rather than user onboarding. That said, the benchmarks are thorough, the integration story is genuinely elegant, and the runtime is self-contained with minimal dependencies. Worth evaluating if your use case fits, but treat it as experimental until the star count climbs.

Sign up to read the full AI review Sign Up Free

Similar repos coming soon.