Summary
Overview
A patch on top of 202610.post1, and nothing about installing or upgrading changes. What changes is that the six things ESPnet does for you outside a training script have one name each and reach you four ways: espnet transcribe, translate, phonemize, align, synthesize and enhance at a terminal, the same six as MCP tools an assistant can call, a notebook each, and a Space each. Two of those are new — espnet phonemize reads the phones in a recording with POWSM, and espnet align lines text up with the audio it was said in, on the CTC head of any model — and they arrive with two new Spaces, powsm-ctc and forced-alignment. Every demo is now built from a directory in this repository, and the two OWSM Spaces stopped carrying their own copy of the demo page: they import the one espnet demo serves, which reads its menus, its window and its tasks off the checkpoint, and 746 lines of app became 166. Alignment likewise has one implementation, and the ctc_segmentation algorithm that was written twice is written once, in espnet2.bin.ctc_segment. Older checkpoints load again: a model published before June 2025 names an initialization this version removed, which no longer stops it. doc/front_ends.md is how the seventh task gets all four front ends, and the table of what runs where is in the README, checked against the code rather than remembered.
Requirements — nothing to do
| 202610.post1 | 202610.post2 | |
|---|---|---|
| Python | >=3.12,<3.14 |
unchanged |
| PyTorch | 2.11.0, 2.13.0, 2.14.0 | unchanged |
pip install espnet |
inference only; training is espnet[train] |
unchanged |
A patch release, so upgrading is pip install -U espnet and nothing else.
espnet demo and the Spaces need pip install espnet[demo] for gradio, and the MCP server pip install espnet[mcp] — neither is a dependency, because no inference path should need a web framework or a protocol server to import.
Important PRs
New Features
Recipe
- PR #6793: A Space for forced alignment (by @sw005320)
- PR #6792: A Space for POWSM-CTC, and its four tasks in the browser (by @sw005320)
- PR #6787: One CTC segmentation, not two — and one naming for the alignment modules (by @sw005320)
- PR #6780: MCP: translate, phonemize and align, under the names the CLI uses (by @sw005320)
- PR #6779: One name a task:
espnet transcribeandespnet synthesize(by @sw005320) - PR #6763: [espnet3] readme -> README in recipe (by @Masao-Someki)
- PR #6761: Follow espnet/notebook's flat Demos layout (by @sw005320)
- PR #6738: [espnet3] Bugfix for espnet3.parallel, local environment (by @Masao-Someki)
Bugfix
- PR #6796: The doc build dies on a page it will not name (by @sw005320)
- PR #6790: Two ways an alignment goes wrong quietly (by @sw005320)
- PR #6782: An initialization this version removed must not stop a model loading (by @sw005320)
- PR #6764: [espnet3] Bugfix - freezing model parameters (by @Masao-Someki)
Documentation
- PR #6794: One page, and one specification for the six tasks (by @sw005320)
- PR #6778: Say which tasks run where (by @sw005320)
- PR #6777: One alignment, on torchaudio's forced alignment (by @sw005320)
- PR #6776: The demo page is the checkpoint's, not OWSM's (by @sw005320)
- PR #6775: Add
espnet phonemize: the phones in a recording, with POWSM (by @sw005320) - PR #6768: Say which demos are checked, where people click them (by @sw005320)
- PR #6749: [Audio metrics 02/12] Add Uni-VERSA regression with checkpoint-compatible parameters (by @ftshijt)
- PR #6747: Support structured annotations and missing references (by @ftshijt)
CI
- PR #6798: A Space that is not uploaded yet is not a broken link (by @sw005320)
- PR #6765: Wait a rate limit out rather than call the demo links broken (by @sw005320)
