modelscope/FunASR
1
0
Fork 0
mirror of https://github.com/modelscope/FunASR.git synced 2026-10-12 02:19:18 +00:00
Code Issues Packages Projects Releases Wiki Activity
5,953 Commits 468 Branches 66 Tags
12e417f4fa4490296ea2d81c5352ea2de8e4cac2
Commit Graph
5953 Commits
This Branch
This Branch
All Branches
Author SHA1 Message Date
zhifu gao 12e417f4fa Merge pull request #3736: correct LiteLLM website transcription route 
Validate the combined website and SenseVoice tree against current main.

Signed-off-by: LauraGPT <[email protected]>
 2026-09-29 14:43:39 +08:00
zhifu gao 15c1959811 Merge pull request #3737: preserve SenseVoice word boundaries 
Normalize leading and standalone SentencePiece boundaries and add CPU timestamp regression coverage.

Signed-off-by: LauraGPT <[email protected]>
 2026-09-29 14:26:04 +08:00
LauraGPT a57cc3f92a fix(sensevoice): keep words separate at standalone boundaries 
Signed-off-by: LauraGPT <[email protected]>
 2026-09-29 06:12:59 +00:00
LauraGPT 2ddabb2c91 fix(sensevoice): normalize first word before joining subpieces 
Signed-off-by: LauraGPT <[email protected]>
 2026-09-29 05:58:08 +00:00
LauraGPT f950854db3 fix(website): correct LiteLLM transcription route and guards 
Signed-off-by: LauraGPT <[email protected]>
 2026-09-29 03:10:14 +00:00
zhifu gao 55203a9995 feat(benchmark): report first nonempty transcript latency (#3734) 
feat(benchmark): report first nonempty transcript latency
 2026-09-28 21:57:23 +08:00
LauraGPT 99ef3b727b feat(benchmark): report first nonempty transcript latency 
Signed-off-by: LauraGPT <[email protected]>
 2026-09-28 13:36:31 +00:00
zhifu gao e1ceba0aa8 Merge pull request #3733 from modelscope/codex/speaker-overlap-accounting-20260928 
fix: assign sentence speakers by total overlap without double counting
 2026-09-28 20:55:07 +08:00
LauraGPT e7006a9315 fix: count speaker overlap once per diarization segment 
Signed-off-by: LauraGPT <[email protected]>
 2026-09-28 12:50:06 +00:00
mehdi BMehdi BoutayebClaude Opus 4.8
 		da12182b6b Fix invalid escape sequences in text normalization and models (#3709) 
Three modules use invalid escapes in non-raw string literals:
- fun_text_processing/text_normalization/normalize.py - \., \w
- funasr/models/fsmn_kws/encoder.py - \[
- funasr/models/fun_asr_nano/tools/format5res.py - \.

On Python 3.12+ each raises SyntaxWarning: invalid escape sequence on
import, and FunASR runs the 3.10-3.12 CI matrix. Escaping the
backslashes leaves every string value identical (ruff W605-style).


Claude-Session: https://claude.ai/code/session_014HBQNzAf5C2E3MiJC48HQw

Co-authored-by: Mehdi Boutayeb <[email protected]>
Co-authored-by: Claude Opus 4.8 <[email protected]>
 2026-09-28 16:08:01 +08:00
Jingguo Yao 2d4566d4a4 Remove unused imports (#3731) 
Removes two unused slice_padding_audio_samples imports.
 2026-09-27 11:17:40 +08:00
YYDH 9e7f6af8c3 fix: avoid deep-copying the checkpoint in load_pretrained_model (#3729) 2026-09-26 18:56:23 +08:00
Bruce-YiiandBruce-Yii 41778c45de fix(nano): correct punctuation timestamps across VAD merges (#3703) 
* fix(nano): correct punctuation timestamps across VAD merges

Anchor zero-acoustic-mass punctuation at the preceding spoken token end (FunASR #3702). Reviewed patch only.

* fix(nano): keep byte-fragment spoken timestamps

* test(nano): cover byte-fragment timestamp preservation

* fix(nano): preserve semantic symbol timestamp spans

* test(nano): preserve semantic symbol timestamp spans

---------

Signed-off-by: Bruce-Yii <[email protected]>
Co-authored-by: Bruce-Yii <[email protected]>
 2026-09-25 10:16:47 +08:00
LiuZi816 3ff9259aad docs: 新增《在昇腾 NPU 上运行 FunASR》指南 (#3708) 
* Create ascend_npu.md

* Update ascend_npu.md

* Update ascend_npu.md

* docs: 补全 funasr 1.3.14 声明依赖清单并区分镜像预装

补全 hydra-core、modelscope、huggingface_hub、safetensors、transformers、tiktoken 等全部声明依赖（取自 PyPI 1.3.14 元数据），注明目标版本；区分常见 NPU 镜像预装与需手动安装的包；标注此前 dry-run 解析出的 7 个包为该环境当次差异、非通用依赖表。回应 #3707 的审查意见。

* docs: 固定 funasr==1.3.14 并限定 torch 配对论述的适用范围

安装命令钉到 1.3.14，与依赖清单及验证环境一致；将已配对 torch 不被解析触碰的结论限定到 #3707 的 dry-run 环境与命令；写明 --no-deps 仅覆盖第一条命令、后续普通 pip 安装仍解析传递依赖，建议安装前 dry-run 检查或用 constraint 固定 torch。
 2026-09-24 10:18:09 +08:00
zhifu gao 02f8b43fc9 docs(nano): clarify training-speed evidence and release scope (#3725) 
docs(nano): clarify training-speed evidence and release scope
 2026-09-21 19:42:39 +08:00
zhifu gao f027d65110 docs(nano): clarify training speed evidence and release scope 
Signed-off-by: zhifu gao <[email protected]>
 2026-09-21 11:36:49 +00:00
zhifu gao 8d8a1a6b6e docs: note mixed-device support in PyPI 1.4.16 (#3724) 
docs: note mixed-device support in PyPI 1.4.16
 2026-09-21 18:34:47 +08:00
zhifu gao 0aaf6652f6 docs: note mixed-device support in PyPI 1.4.16 
Signed-off-by: zhifu gao <[email protected]>
 2026-09-21 10:16:28 +00:00
li-lizhe bad0781a2e fix: use y.device instead of hardcoded .cuda() in data2vec compute_var (#3710) 
* fix: use y.device instead of hardcoded .cuda() in data2vec compute_var

In compute_var(), zc = torch.tensor(y.size(0)).cuda() unconditionally
calls .cuda() which raises RuntimeError: Torch not compiled with CUDA
enabled on non-CUDA accelerators (Ascend NPU, XPU, HPU, MPS).

Use device=y.device to respect the input tensor's current device.

Signed-off-by: li-lizhe <[email protected]>

* test: add CPU/Gloo regression tests for data2vec compute_var

Cover both the uninitialized (no process group) branch and a real
2-rank Gloo process group with unequal local row counts, pinning the
device-agnostic contract introduced in the parent commit (PR #3710).

Signed-off-by: li-lizhe <[email protected]>

* fix(test): resolve compute_var import and f32_worker spawn pickle

* fix(test): ensure Manager and workers are cleaned up on failure paths

join(timeout) does not terminate a still-running worker, so a crashed or
hung rank could leak a subprocess on test failure. Wrap the two-rank
runner in try/finally that terminate()+join()s any alive worker and shuts
down the Manager.

Signed-off-by: li-lizhe <[email protected]>

---------

Signed-off-by: li-lizhe <[email protected]>
 2026-09-21 10:28:58 +08:00
Hexu Zhao 782a7a99d5 [perf] Fun-ASR-Nano fine-tuning on B200: 0.244 s to 0.068 s per step at 1 GPU, mostly from cuDNN SDPA plan builds on every new batch shape (#3705) 
* fun_asr_nano: opt-in scoped SDPA backends and compiled decoder for fine-tuning

* tests: FunASRNano opt-ins leave defaults untouched

* fun_asr_nano: declare the torch 2.3 floor the fine-tuning recipe needs
 2026-09-21 10:20:05 +08:00
Xu Xiang 1878d61dbe fix(paraformer): correct timestamp alpha and peak argument order (#3716) 
Pass CIF alpha weights and peak indices to the timestamp helper in its declared order and add boundary/content regressions. Fixes #3723.
 2026-09-19 10:42:09 +08:00
Gmgge b096b7950f fix(campplus): bound clustering thread usage and spectrum work (#3699) 
Apply ncpu to BLAS without rolling back torch OpenMP settings, compute only the required leading spectral pairs, and add regression coverage for repeated AutoModel construction and process-wide thread limits.
 2026-09-19 10:32:10 +08:00
zhifu gao cf33286387 Record FunASR 1.4.16 release closeout (#3722) 
Record the verified release, package, website deployment, announcement, and model-card audit in the community growth ledger.

Signed-off-by: LauraGPT <[email protected]>
 2026-09-18 15:41:06 +08:00
zhifu gao acca7ea070 Publish FunASR 1.4.16 bilingual site guide (#3721) 
Publish bilingual 1.4.16 release pages, catalogue and sitemap entries, and exact release-asset contracts.

Signed-off-by: LauraGPT <[email protected]>
 2026-09-18 14:58:31 +08:00
zhifu gao 904cd18681 Release FunASR 1.4.16 (#3719) 
Prepare the 1.4.16 Python release, synchronize four-language release entry points, and make the native Transformers site contract follow the current package version.

Signed-off-by: LauraGPT <[email protected]>
v1.4.16 		2026-09-18 12:03:34 +08:00
zhifu gao 615ef059c9 Merge pull request #3714 from modelscope/codex/mixed-device-availability-20260916 
docs: clarify mixed-device source version requirement
 2026-09-16 12:55:18 +08:00
zhifu gao b2d63f021a docs: clarify mixed-device source version requirement 
Signed-off-by: zhifu gao <[email protected]>
 2026-09-16 04:21:01 +00:00
zhifu gao 4d8e08748c Merge pull request #3706 from damnwenxi/fix/submodel-device-placement 
fix: honor submodel device placement across inference calls
 2026-09-16 10:27:29 +08:00
LauraGPT 1605ad20a3 test: isolate optional clustering dependency 
Signed-off-by: LauraGPT <[email protected]>
 2026-09-16 02:21:55 +00:00
zhifu gao 6444e200ca Merge pull request #3713 from modelscope/codex/growth-target-strict-20260916 
docs: correct strict ecosystem growth target
 2026-09-16 08:11:58 +08:00
zhifu gao fc98a60342 docs: correct strict ecosystem growth target 
Signed-off-by: zhifu gao <[email protected]>
 2026-09-15 23:34:51 +00:00
zhifu gao 66eb99a3d2 Merge pull request #3712 from modelscope/codex/roadmap-status-20260916 
docs: align bilingual roadmap with delivered milestones
 2026-09-16 02:24:35 +08:00
zhifu gao 8181a0225e docs: align bilingual roadmap with delivered milestones 
Signed-off-by: zhifu gao <[email protected]>
 2026-09-15 17:47:50 +00:00
zhifu gao b07c5facc8 Merge pull request #3711 from modelscope/codex/sensevoice-punctuation-doc-20260915 
docs(punctuation): clarify SenseVoice postprocessing
 2026-09-15 23:14:39 +08:00
zhifu gao 42a7f02d93 test(punctuation): guard SenseVoice documentation on exact PR heads 
Signed-off-by: zhifu gao <[email protected]>
 2026-09-15 14:30:57 +00:00
zhifu gao 176e35ec7c docs(punctuation): clarify SenseVoice postprocessing 
Signed-off-by: zhifu gao <[email protected]>
 2026-09-15 14:26:43 +00:00
damnwenxi eee63614f0 test: avoid Python 3.11-only unittest cleanup API 2026-09-14 18:42:44 +08:00
damnwenxi aa353d6d22 fix: honor submodel device placement across inference calls 2026-09-14 09:20:40 +08:00
zhifu gao 486b4b7ceb docs: connect native Transformers guides to tested CUDA recipe (#3697) 
Connect bilingual source guides to the tested canonical GPU example without changing CPU recipes or model capabilities.

Signed-off-by: LauraGPT <[email protected]>
 2026-09-10 18:58:26 +08:00
zhifu gao 2f0de0ef52 Merge pull request #3696 from modelscope/codex/pages-sdk-entry-20260910 
docs: connect Pages API references to practical guides
 2026-09-10 17:33:30 +08:00
zhifu gao 87bcc4bcc5 docs: connect Pages API references to practical guides 
Signed-off-by: zhifu gao <[email protected]>
 2026-09-10 09:18:28 +00:00
zhifu gao c87086f1ce Merge pull request #3695 from modelscope/codex/sensevoice-alignment-docs-20260910 
docs: clarify SenseVoice word timestamp alignment
 2026-09-10 16:51:06 +08:00
zhifu gao a111c75059 docs: clarify SenseVoice word timestamp alignment 
Signed-off-by: zhifu gao <[email protected]>
 2026-09-10 08:33:18 +00:00
zhifu gao 85516946f7 docs: distinguish native Transformers model downloads (#3694) 
Keep native and toolkit artifacts distinct across four-language model entry points.

Signed-off-by: zhifu gao <[email protected]>
 2026-09-10 14:54:55 +08:00
zhifu gao 1d2307b1ab docs: distinguish native Transformers model downloads 
Signed-off-by: zhifu gao <[email protected]>
 2026-09-10 06:39:10 +00:00
zhifu gao 535e0d9d34 docs: ground punctuation tutorial in observed CPU output (#3693) 
Preserve original text, observed candidates and manual-review boundaries.

Signed-off-by: zhifu gao <[email protected]>
 2026-09-10 14:06:16 +08:00
zhifu gao f98502d8a4 docs: ground punctuation tutorial in observed CPU output 
Signed-off-by: zhifu gao <[email protected]>
 2026-09-10 05:44:35 +00:00
zhifu gao 83a3e3dc9c Merge pull request #3692 from modelscope/codex/subtitle-output-guide-20260910 
docs: clarify SRT and WebVTT subtitle export workflow
 2026-09-10 11:59:01 +08:00
zhifu gao dfbda65aad docs: make subtitle export examples reproducible and bounded 
Signed-off-by: zhifu gao <[email protected]>
 2026-09-10 03:43:33 +00:00
zhifu gao 9fbb8a8832 Merge pull request #3691 from modelscope/codex/sensevoice-tags-story-20260910 
docs: make SenseVoice emotion-tag guidance reproducible and bounded
 2026-09-10 10:21:10 +08:00