lilfugu-experimental

Aggressive variant of lilfugu with stronger term conversion.

This variant converts terms more aggressively — higher benchmark scores, but may over-convert in some cases. For a more conservative approach, see lilfugu.

Comparison with lilfugu

	lilfugu	lilfugu-experimental
Term conversion	Conservative	Aggressive
DevTerm Composite	0.6272	0.7648
JSUT CER (general)	10.8%	11.9%

Benchmarks

ADLIB

Model	CER	Term Accuracy (Exact)	Composite
lilfugu-experimental	14.6%	68.5%	0.7648
lilfugu	26.3%	51.6%	0.6272
Qwen3-ASR-1.7B (base)	41.1%	24.6%	0.4203
Whisper large-v3-turbo	41.9%	20.2%	0.3935

Benchmark: ADLIB — Language-aware ASR benchmark for Japanese

JSUT

Model	CER
Qwen3-ASR-1.7B (base)	10.7%
lilfugu	10.8%
lilfugu-experimental	11.9%

Dataset: JSUT

Variants

Repository	Size	Format
lilfugu-experimental (this)	4.1 GB	MLX bfloat16
lilfugu-experimental-8bit	2.8 GB	MLX 8bit quantized
lilfugu-experimental-transformers	4.1 GB	safetensors fp16 (CUDA/Linux)
lilfugu-experimental-transformers-8bit	2.2 GB	bitsandbytes int8 (CUDA/Linux)
lilfugu-experimental-lora	~49 MB	LoRA adapter

Usage

MLX (Apple Silicon)

pip install -U mlx-audio

from mlx_audio.stt import load

model = load("holotherapper/lilfugu-experimental")
result = model.generate("audio.wav", language="Japanese")
print(result.text)

CUDA / Linux

from qwen_asr import Qwen3ASRModel

model = Qwen3ASRModel.from_pretrained("holotherapper/lilfugu-experimental-transformers")
result = model.transcribe("audio.wav")

License

Apache 2.0