ULUNAS Speech Enhancement

ULUNAS is an ultra-lightweight U-Net for real-time speech enhancement. This package provides an FP32 LiteRT model for removing background noise from 16 kHz mono audio samples.

✨ Key Highlights

• Real-time performance — RTFx 23.14 (over 23× faster than real time) on an Android™ Vivo X300 using one Arm® CPU thread.
• Compact model — 5.71 MB.

📦 Model Details

Model Description

ULUNAS uses an ultra-lightweight U-Net selected through neural architecture search. Its upstream architecture combines efficient convolutional blocks with affine PReLU activations and causal time-frequency attention for speech enhancement on low-footprint devices.

• Developed by: Xiaobin Rong, Leyan Yang, Dahan Wang, Yuxiang Hu, Changbao Zhu, Kai Chen, and Jing Lu
• Model type: Streaming U-Net speech-enhancement model
• License: MIT
• Base model: ULUNAS trained on DNS3
• Packaged variant: FP32 LiteRT model
• Package form: LiteRT model, runtime manifest, representative input and output, and checksums

Model Sources

• Base model: ULUNAS DNS3 checkpoint at revision 631d2c8
• Upstream repository: Xiaobin-Rong/ul-unas
• Paper: UL-UNAS: Ultra-Lightweight U-Nets for Real-Time Speech Enhancement via Network Architecture Search

🚀 Get Started with the Model

🔓 Compute Flow — Early Access

The inference engine for this model package is available through the Compute Flow Early Access Program.

Want to try it?

📩 Contact us at ai-early-access@arm.com to request access.

📊 Quality evaluation

Accuracy has not been evaluated. The model uses the FP32 data type, so its accuracy is expected to match the original implementation.

🎯 Performance evaluation

Performance was measured under the following conditions:

• Target: Android™ Vivo X300 smartphone.
• CPU: One Arm® CPU thread.
• Runs: 5 warmups followed by 30 measured fresh-process runs, with one second between runs.

The following methodology and definitions were used:

• Runtime: LiteRT 2.1.6 with the XNNPACK CPU backend.
• Input: The distributed 10-second, 16 kHz mono samples/noisy.wav sample.
• Latency is the profiled audio-enhancement run() scope, including 626 streaming LiteRT frame invocations and caller overhead. Model setup, warmup, WAV decoding, and output file I/O are excluded.
• Real-time factor (RTF) is latency divided by the 10-second input duration. Values below 1 are faster than real time, and lower values are faster. Percentiles are calculated from the 30 measured runs.

⚠️ Known Limitations

• The packaged contract supports 16 kHz mono audio only.

🛠️ Technical Specifications

Objective

Enhance noisy speech as a stream while keeping model execution state in the runtime and exposing a fixed-shape FP32 frame contract to the caller.

Runtime Architecture

Precision and Quantization

The model artifact, inputs, outputs, and streaming states use FP32. No quantization is applied.

Input Specification

The input is a 16 kHz mono audio sample. Frames advance by the manifest's 256-sample hop length.

Output Specification

enhanced_frame is an FP32 tensor with shape [1, 512]. The caller overlap-adds and normalizes the enhanced frames to produce the output audio sample.

Repository Contents

• benchmarks/ulunas-speech-enhancement-vivo-x300-baseline.yaml — Android™ Vivo X300 performance report.
• manifest.json — runtime model, tensor, state, and audio-configuration contract.
• metadata.yaml — model identity, origin, format, precision, and publication metadata.
• samples/enhanced.wav — 16 kHz mono FP32 output generated from samples/noisy.wav with the packaged artifact.
• samples/noisy.wav — representative 10-second, 16 kHz mono PCM input from the upstream project.
• ulunas_fused_frame_fp32_litert.tflite — FP32 LiteRT streaming speech-enhancement model.
• SHA256SUMS — model-package checksums for reproducibility.

🗂️ Model and Asset Origin

• ulunas_fused_frame_fp32_litert.tflite was re-exported from the upstream model_trained_on_dns3.tar PyTorch checkpoint. The conversion applies an exact current-frame rewrite for decoder transposed convolutions and includes the fixed frame transforms in the graph.
• samples/noisy.wav is byte-identical to upstream audio/noisy/0119.wav.
• samples/enhanced.wav was generated from samples/noisy.wav with the packaged LiteRT artifact.
• manifest.json was authored for this package from the validated LiteRT tensor and state contract.
• metadata.yaml records the packaged model identity and publication fields.
• benchmarks/ulunas-speech-enhancement-vivo-x300-baseline.yaml records the performance results presented in this model card.

🔐 Checksums

SHA256SUMS was generated by recursively hashing every regular file in the model bundle, including files in subdirectories, except the generated root SHA256SUMS and paths with a dotfile component.

From the model bundle root, verify the checked-out files with:

shasum -a 256 -c SHA256SUMS

Citation

@ARTICLE{ulunas,
  author={Rong, Xiaobin and Yang, Leyan and Wang, Dahan and Hu, Yuxiang and
    Zhu, Changbao and Chen, Kai and Lu, Jing},
  journal={IEEE Transactions on Audio, Speech and Language Processing},
  title={UL-UNAS: Ultra-Lightweight U-Nets for Real-Time Speech Enhancement via
    Network Architecture Search},
  year={2026},
  volume={34},
  number={},
  pages={1085-1096},
  keywords={Computational modeling;Computer architecture;Convolution;Speech enhancement;
    Computational efficiency;Performance evaluation;Computational complexity;Artificial
    intelligence;Time-frequency analysis;Time-domain analysis;Speech enhancement;
    ultra-lightweight;neural architecture search;computational complexity},
  doi={10.1109/TASLPRO.2026.3661271}}