DiTReducio: Training-Free Calibration and Acceleration for DiT-based Text-to-Speech

Official implementation of DiTReducio, a training-free calibration and acceleration framework for DiT-based text-to-speech models. DiTReducio identifies temporal and branch redundancy in DiT inference and applies compression strategies via progressive calibration to achieve speedup without training cost.

News

• 2026/04: 🎉 Our paper has been accepted to ACL 2026 Findings!

Overview

DiTReducio is a training-free acceleration framework that eliminates redundant computations in DiT-based TTS models via progressive calibration.

Core Compression Strategies

• Temporal Skipping (TS): Caches module outputs at a given timestep and reuses them in subsequent steps to avoid temporally redundant computation.
• Branch Skipping (BS): Skips the redundant unconditional branch in Classifier-Free Guidance and reconstructs it via a Branch Residual mechanism to preserve essential guidance details.

Three-Phase Progressive Calibration

1. Check Phase: Identifies highly temporally redundant layer-step pairs by detecting diagonal-like attention patterns.
2. Pre-Calibration Phase: Selectively applies TS to marked pairs to ensure a superior strategy combination and avoid suboptimal compression.
3. Calibration Phase: Systematically applies both TS and BS across all layer-step pairs.

Results

Performance on LibriSpeech-PC-test-clean (averaged over 5 seeds):

Model	Metric	T0 (Baseline)	T1	T2	T3	T4	T5	T6
F5-TTS	SIM-o	0.640	0.640	0.637	0.629	0.618	0.610	0.590
	WER (%)	2.636	2.655	2.564	2.643	2.634	2.661	2.900
	RTF	0.178	0.165	0.149	0.138	0.129	0.120	0.112
	Ops Ratio (%)	100.00	82.59	66.38	55.09	45.58	39.26	34.42
MegaTTS 3	SIM-o	0.750	0.750	0.748	0.743	0.734	0.691	0.626
	WER (%)	3.112	3.112	3.110	3.073	3.095	3.133	3.030
	RTF	0.396	0.395	0.359	0.287	0.224	0.176	0.156
	Ops Ratio (%)	100.00	98.87	88.02	68.19	48.94	33.88	27.52

T4 represents the optimal balance point. DiTReducio achieves 1.37x speedup for F5-TTS and 1.76x for MegaTTS 3 at T4 with no significant quality degradation.

Installation

cd DiTReducio
uv venv && source .venv/bin/activate
uv pip install -e .

Backend Dependencies

DiTReducio requires the upstream TTS model code:

• F5-TTS: Clone from SWivid/F5-TTS and set the path in config
• MegaTTS 3: Clone from bytedance/MegaTTS3 and set the path in config

Or use the provided setup script:

bash scripts/fetch_backends.sh <target-root>

Quick Start

1. Configure

Copy and edit the example config:

cp configs/f5tts.example.yaml configs/local.f5tts.yaml
# Edit paths in configs/local.f5tts.yaml

Key path fields:

Field	Description
paths.backend_code_root	F5-TTS or MegaTTS 3 code root directory
paths.backend_ckpt_root	Model weights directory
paths.strategy_output_root	Strategy table output directory
paths.inference_output_root	Inference audio output directory

2. Calibrate

Run the three-phase calibration to generate a strategy table:

# F5-TTS
python -m ditreducio.cli.calibrate --backend f5tts --config configs/local.f5tts.yaml --delta 0.2

# MegaTTS 3
python -m ditreducio.cli.calibrate --backend megatts3 --config configs/local.megatts3.yaml --delta 0.8

3. Accelerated Inference

Load the saved strategy table and run accelerated inference:

# F5-TTS
python -m ditreducio.cli.infer --backend f5tts --config configs/local.f5tts.yaml --delta 0.2

# MegaTTS 3
python -m ditreducio.cli.infer --backend megatts3 --config configs/local.megatts3.yaml --delta 0.8

Experiments

Threshold Sweep (T0–T6)

# Full sweep: calibrate + infer + eval for all thresholds
python scripts/run_sweep_f5.py \
    --backend_root /path/to/F5-TTS \
    --f5tts_ckpt /path/to/model_1250000.safetensors \
    --vocoder_path /path/to/vocos-mel-24khz \
    --data_root /path/to/LibriSpeech \
    --dataset clean

# Custom thresholds
python scripts/run_sweep_f5.py \
    --backend_root /path/to/F5-TTS \
    --f5tts_ckpt /path/to/model_1250000.safetensors \
    --vocoder_path /path/to/vocos-mel-24khz \
    --data_root /path/to/LibriSpeech \
    --deltas 0.0 0.05 0.1 0.15 0.2 0.25 0.3

Evaluation (WER + SIM-o)

# Auto-download evaluation models
python scripts/eval_metrics.py \
    --gen_dir <dir> --lst_file <lst> --librispeech_root <root> --device cuda

# Use local evaluation model weights
python scripts/eval_metrics.py \
    --gen_dir <dir> --lst_file <lst> --librispeech_root <root> \
    --whisper_ckpt /path/to/faster-whisper-large-v3 \
    --ecapa_ckpt /path/to/wavlm_large_finetune.pth \
    --device cuda

Citation

@article{huo2025ditreducio,
  title={Ditreducio: A training-free acceleration for dit-based tts via progressive calibration},
  author={Huo, Yanru and Jiang, Ziyue and Tang, Zuoli and Hong, Qingyang and Zhao, Zhou},
  journal={arXiv preprint arXiv:2509.09748},
  year={2025}
}

Acknowledgements

Our approach is inspired by DiTFastAttn for training-free acceleration in diffusion transformers. We build upon F5-TTS and MegaTTS3 for the baseline TTS models.