Stream.FM¶

Development line: project:stream-fm · thread stream-fm-development
Last event: 2026-05-14 · 1 dated since 2026-05-14 · Researched: 2026-09-05 · confidence: medium

What it is¶

Stream.FM is an AGPL-3.0 PyTorch research codebase for audio ML engineers building frame-by-frame speech restoration.

• Speech enhancement and dereverberation
• Codec post-filtering, bandwidth extension and STFT phase retrieval
• Mel vocoding through the bundled related MelFlow implementation

Development line¶

• 2026-05-14 — Stream.FM project resources were linked. On 2026-05-14, Stream.FM was linked through an examples site, a GitHub source repository, and a Google Drive folder. Together, these references establish a public project-resource bundle for the line, although they do not establish a versioned release or the contents of the supplementary materials.

What changed¶

• 2025-09-18 — MelFlow, the related streaming Mel-vocoder work later bundled with Stream.FM, appeared as an arXiv preprint.
• 2025-12-22 — Stream.FM appeared as an arXiv preprint.
• 2026-04-21 — arXiv v3 became the latest listed Stream.FM paper revision.
• 2026-05-05 — the authors demonstrated joint real-time speech enhancement and bandwidth extension at ICASSP.
• 2026-05-11 — the initial public Stream.FM and MelFlow source code was committed.
• 2026-05-13 — AGPL-3.0 was added to the source repository.
• 2026-05-14 — the project page, repository and checkpoint-folder links were recorded together; no separate upstream commit or GitHub Release is evidenced for this date.
• 2026-05-23 — repository citation maintenance changed documentation only.
• 2026-06-05 — the last shown upstream push updated the Stream.FM TASL citation only.

How to use this¶

Check the examples site, repository, and Drive materials together when evaluating Stream.FM as of 2026-05-14. No specific setup, result, or release version is verified yet.

1. Create a Python 3.10+ environment and install the documented requirements with a PyTorch CUDA wheel index matching the local CUDA runtime. — https://github.com/sp-uhh/streamfm/blob/main/README.md
2. Download a checkpoint from the documented Google Drive folder into a local checkpoints directory using the maintained gdown command. — https://drive.google.com/drive/folders/1u2QKjGAdxblQVV8-qmifSM9AwhT86LER
3. Choose the task configuration that matches the checkpoint architecture, then run inference.py with an input WAV directory, output directory, solver and checkpoint path. — https://github.com/sp-uhh/streamfm/blob/main/README.md
4. For live frame-by-frame integration, implement state initialization and forward_step calls around the CausalNCSNpp backbone rather than expecting a ready-made streaming server. — https://github.com/sp-uhh/streamfm/blob/main/README.md

Best practices¶

• Start with the authors' suggested 5xeuler solver, and only increase GPU parallelism when the target machine has those GPUs. — https://github.com/sp-uhh/streamfm/blob/main/README.md
• Keep the Hydra config and checkpoint architecture matched; the documented inference path requires this pairing. — https://github.com/sp-uhh/streamfm/blob/main/README.md
• For streaming, evaluate CUDA graphs and the default torch.compile wrapper on the target hardware rather than assuming the paper latency transfers unchanged. — https://github.com/sp-uhh/streamfm/blob/main/README.md
• Benchmark end-to-end target-hardware latency; the published 48 ms result is a measured research setup, not a hardware-independent guarantee. — https://arxiv.org/abs/2512.19442
• Load only trusted checkpoints: inference uses torch.load with weights_only=False. — https://raw.githubusercontent.com/sp-uhh/streamfm/main/inference.py
• Review AGPL-3.0 obligations before incorporating repository code into a product or network service. — https://github.com/sp-uhh/streamfm/blob/main/LICENSE

Superseded by this¶

• 2026-05-14 — Corrected: this is a date on which the links were recorded, not a separately evidenced upstream Stream.FM release; public source code began on 2026-05-11 and GitHub lists no releases.
• 2026-05-23 — Citation-only repository maintenance is not evidence of a new model, checkpoint or runtime release.
• 2026-06-05 — The TASL citation update is documentation-only; no later official model, dataset or package release was found in the checked sources.

Still unknown¶

• The linked Drive folder could not be enumerated here, so its 2026-05-14 checkpoint inventory, hashes, availability and contents are unverified.
• AGPL-3.0 is verified for the repository source; no separate checkpoint licence or model card for the Drive-hosted weights was verified.
• The paper's latency results do not prove target-device, audio-I/O or production latency.
• No GitHub Release, package artifact or checkpoint-version manifest was found; adopters need to record the exact source commit and checkpoint identity they use.

Sources¶

Agent brief¶

• Subject: project:stream-fm, thread stream-fm-development, 1 dated events 2026-05-14 → 2026-05-14.
• Practical note: As of 2026-05-14, practitioners evaluating Stream.FM should consult the examples site, source repository, and supplementary Drive materials together; no specific setup, result, or release version is yet verified.
• Confidence: medium. Dated supersedes above are the authority for what is obsolete.