DPDFNet

DPDFNet is a free, open-source model for reducing noise in single-channel speech, with support for recordings, live streams and edge-device applications. It builds on DeepFilterNet2 using dual-path encoder blocks and an always-on fine-tuning approach intended to reduce over-attenuation. Pretrained versions support audio at 8, 16 and 48 kHz. The project provides PyTorch code and ONNX and TFLite inference models, plus command-line, Python and stateful streaming interfaces. In streaming mode, the first enhanced output follows about 20 ms of buffering; later blocks add about 10 ms of delay. A hosted Gradio demo and audio examples are linked from the project. Listed integrations include sherpa-onnx, OBS Studio, VST3 and LADSPA plugins, Core ML ports and WebAssembly. The project page reports real-time deployment of DPDFNet-4 on Ceva-NeuPro-Nano edge NPUs in a benchmark setup using int8 weights and int16 activations. Code is under Apache License 2.0. Model files are downloaded separately using repository instructions for checkpoints and ONNX or TFLite models.

DPDFNet offers multiple deployment formats and interfaces for single-channel speech enhancement. Plan to download model files separately, and account for the documented buffering in streaming use.

DPDFNet suits developers and teams building speech enhancement into recording, streaming or edge-device workflows. Its multiple inference formats and documented interfaces offer several ways to integrate it.

DPDFNet is a free, open-source model for enhancing noisy, single-channel speech in recordings, live streams, and edge devices. It is aimed at developers and audio teams that need control over runtimes and deployment, rather than a finished consumer editor. Its broad inference options and streaming interfaces are compelling, provided you are comfortable managing model downloads and integration work.

Overview

DPDFNet extends DeepFilterNet2 with dual-path encoder blocks and an always-on fine-tuning method intended to limit over-attenuation. In practical terms, that is a design focused on removing noise without needlessly thinning the voice, a useful priority for speech archives, broadcasts, and embedded products. It remains a specialized enhancement model, not a general audio-production workspace.

Pretrained variants cover 8, 16, and 48 kHz audio, so teams can match common narrowband, wideband, or high-quality pipelines. The Apache License 2.0 permits use and modification under its terms, making the project suitable for products that need an auditable, adaptable foundation.

Key features

The project supplies PyTorch code alongside ONNX and TFLite inference models. That range can simplify fitting the enhancer into research code, mobile or embedded runtimes, and other deployment stacks without rewriting the model for a single framework. The command-line interface is practical for scripted processing; the Python API suits application integration; and the stateful streaming API addresses continuous audio rather than isolated files.

Streaming behavior is documented precisely: the first enhanced output follows about 20 ms of buffering, with roughly 10 ms of additional delay for subsequent blocks. Those figures give engineers a concrete starting point for latency budgeting, while applications with extremely tight timing requirements will still need to assess the impact in their own pipeline.

A hosted Gradio demo and project audio examples provide a quick way to hear the intended enhancement. Community integrations include sherpa-onnx, OBS Studio, VST3 and LADSPA plugins, Core ML ports, and WebAssembly, which broadens the paths from model to working application. The project page also reports real-time deployment of DPDFNet-4 on Ceva-NeuPro-Nano edge NPUs using int8 weights and int16 activations in its benchmark setup. That demonstrates an edge-deployment route, but does not promise identical performance on other hardware.

Evaluation described in the paper uses long, low-SNR recordings in 12 languages and everyday noise conditions. Model files are not included in the source repository: checkpoints and ONNX or TFLite packages must be downloaded from Hugging Face, adding a setup step before inference.

Pricing

Open-source DPDFNet — free (Apache-2.0 licensed code and models; no commercial pricing or usage limits stated).

This is the right plan for developers, researchers, and product teams able to host and operate their own processing. There are no paid seats or service quotas to budget for, but the project does not provide a turnkey hosted workflow; downloading models and building the surrounding application remain your responsibility.

Platforms

DPDFNet supports Android, iOS, Linux, macOS, Windows, API use, and self-hosted deployment. PyTorch, ONNX, and TFLite formats, together with the documented integrations, make it a strong fit when one model must travel across desktop, mobile, server, and embedded environments. Its processing mode covers both batch and streaming scenarios, with audio as the input type.

Who it's for

Choose DPDFNet if you are integrating speech cleanup into software, operating a self-hosted audio service, or targeting an edge device and want permissively licensed code with multiple inference formats. It is especially appropriate for teams that can handle model acquisition, runtime configuration, and latency decisions themselves.

It is less suitable for someone who simply wants to upload a clip and download a finished result. The separate model download and engineering-oriented interfaces add friction compared with consumer web or mobile tools.

Pros and cons

Pros

• Flexible deployment: PyTorch, ONNX, and TFLite models support varied runtime strategies.
• Useful interfaces: CLI, Python, and stateful streaming APIs cover scripts, applications, and live pipelines.
• Clear latency expectations: approximately 20 ms initial buffering and 10 ms subsequent block delay help with architecture planning.
• Broad integration path: community options span OBS, plugin formats, Core ML, WebAssembly, and sherpa-onnx.
• Permissive cost structure: Apache 2.0 licensing and a free plan avoid recurring service charges.

Cons

• Extra setup: checkpoints and inference models must be downloaded separately from Hugging Face.
• Engineering required: the project provides model interfaces rather than a polished consumer editing workflow.
• Streaming delay remains: buffering and block latency may conflict with the strictest real-time applications.

Alternatives

LALAL.AI is a freemium choice across mobile, desktop, web, and API platforms, with a free starter queue and previews; choose it when you prefer a hosted workflow over integrating an open model.

Denoise targets Android and iOS with a free-to-try tier and a paid Pro plan, making it a simpler mobile-first option.

Riverside Magic Audio combines freemium audio enhancement with recording features on Android, iOS, macOS, and web, suited to creators working inside Riverside.

AudioEnhancerMAX is another free option spanning Android, Linux, macOS, web, and Windows.

Agora AI Noise Suppression uses usage-based pricing at $0.59 per 1,000 minutes, with the first 10,000 minutes free monthly, for teams wanting a managed API model.

Krisp AI Meeting Assistant is a paid meeting-focused product with a seven-day free trial rather than an open deployment model.

Auphonic offers freemium, API-accessible production with two free hours per month, fitting automated post-production workflows.

Adobe Enhance Speech is a web-only freemium tool with per-file and daily limits, best for straightforward browser-based enhancement.

Browse the AI Speech Enhancement Software list for more options.

Verdict

DPDFNet is a strong recommendation for technically capable teams that need a free, Apache-licensed speech-enhancement model deployable through several runtimes and streaming interfaces.