Open-source speech recognition toolkit for training, inference, streaming ASR, VAD, punctuation, speaker diarization pipelines, and OpenAI-compatible/MCP serving.
Python asraudiochineseemotion-recognitionfunasrmcp-servermultilingual-asropenai-compatible-apiparaformerpunctuationpytorchreal-time-asrspeaker-diarizationspeech-recognitionspeech-to-textstreaming-asrtranscriptionvllmvoice-activity-detectionwhisper-alternative
docs: align bilingual roadmap with delivered milestones
zhifu gao8181a0225e00
diff --git a/.github/workflows/test-repository-roadmap-docs.yml b/.github/workflows/test-repository-roadmap-docs.ymlnew file mode 100644index 000000000..f9a2de419--- /dev/null+++ b/.github/workflows/test-repository-roadmap-docs.yml@@ -0,0 +1,33 @@+name: Validate repository roadmap++on:+ pull_request:+ paths:+ - 'docs/repository_roles.md'+ - 'docs/repository_roles_zh.md'+ - 'tests/test_repository_roadmap_docs.py'+ - '.github/workflows/test-repository-roadmap-docs.yml'+ push:+ branches: [main]+ paths:+ - 'docs/repository_roles.md'+ - 'docs/repository_roles_zh.md'+ - 'tests/test_repository_roadmap_docs.py'+ - '.github/workflows/test-repository-roadmap-docs.yml'++permissions:+ contents: read++jobs:+ roadmap-contract:+ runs-on: ubuntu-latest+ timeout-minutes: 5+ steps:+ - uses: actions/checkout@v4+ with:+ ref: ${{ github.event.pull_request.head.sha || github.sha }}+ persist-credentials: false+ - uses: actions/setup-python@v5+ with:+ python-version: '3.12'+ - run: python tests/test_repository_roadmap_docs.py -vdiff --git a/docs/repository_roles.md b/docs/repository_roles.mdindex 5f18c0aa6..b78666ff8 100644--- a/docs/repository_roles.md+++ b/docs/repository_roles.md@@ -93,7 +93,7 @@ The roadmap is a queue of testable outcomes, not a list reserved for maintainers| [Realtime preview efficiency on L20-class GPUs](https://github.com/modelscope/FunASR/issues/3528) | After matching the number of partial messages, which refresh interval and partial window provide the best latency/throughput trade-off without silently skipping previews? | Client JSONL and `--log-decode-profile` server logs from the exact commit, with SPK, ping, audio, concurrency, partial window, and partial-message count held constant | Reproduction on L20, L4, A10, or other non-H100 GPUs; analysis of queue, encoder, and engine time || [AMD Windows Vulkan stability](https://github.com/modelscope/FunASR/issues/3479) | Does the current runtime reach model initialization and transcription on the reporter's AMD GPU, and where is the last successful initialization boundary if it does not? | Exact archive name and SHA256, GPU/driver/Windows versions, full initialization log, and a reporter hardware retest | AMD Windows hardware owners and Vulkan/llama.cpp contributors || [Complete public checkpoint functionality](https://github.com/modelscope/FunASR/issues/3496) | How should the missing CTC tensors be published from an authorized model-owner account and validated after upload? | Immutable model revision, file hashes, public clean-cache download, and real timestamp/diarization inference | Model owners with Hugging Face write access and checkpoint validation experience |-| [Upstream model integrations](https://github.com/huggingface/transformers/pull/46180) | Can Fun-ASR-Nano remain compatible with upstream Transformers while preserving pinned model-card and regression-test boundaries? | Exact-head upstream CI, focused local tests, model-card review, and maintainer review | Transformers reviewers and users who can test downstream loading before merge |+| [Native Transformers compatibility](https://github.com/QwenAudio/Fun-ASR) | Can the merged Fun-ASR-Nano integration retain downstream loading and transcription compatibility across supported Transformers versions? | Exact model and library revisions, focused transcription tests, and model-card capability boundaries | Transformers users who can reproduce downstream compatibility failures |### Before claiming an item@@ -112,19 +112,20 @@ Contributions and issue evidence may be written in Chinese or English. A roadmap### Delivered+- **Fun-ASR-Nano native Transformers integration** - [huggingface/transformers#46180](https://github.com/huggingface/transformers/pull/46180) merged on 2026-09-09. Use the [native `-hf` checkpoint](https://huggingface.co/FunAudioLLM/Fun-ASR-Nano-2512-hf) and the [model repository's Transformers entry point](https://github.com/QwenAudio/Fun-ASR). The native generation path does not include the toolkit's auxiliary CTC branch; this milestone does not establish native timestamps, diarization, or realtime sessions.+- **Qwen3-ASR offline vLLM workflow** - [#3592](https://github.com/modelscope/FunASR/pull/3592) delivered the native `Qwen3ASRModel.LLM` example. The reporter confirmed completion and closed [#3419](https://github.com/modelscope/FunASR/issues/3419) on 2026-09-09. Its accuracy comparison used that report's audio, reference, and custom scorer, not a general CER guarantee.- **Bounded realtime long-session state** — fixes merged via [#3214](https://github.com/modelscope/FunASR/pull/3214) and [QwenAudio/Fun-ASR#135](https://github.com/QwenAudio/Fun-ASR/pull/135), diagnostics shipped, and reporter evidence allowed [#3101](https://github.com/modelscope/FunASR/issues/3101) to close.- **Stable application-facing APIs** — the toolkit now ships an OpenAI-compatible transcription server, health checks, browser and command-line smoke tests, and documented Python / CLI / HTTP / WebSocket entry points in the [deployment matrix](./deployment_matrix.md).- **Industrial and edge deployment paths** — vLLM serving and signed release workflows are documented; the verified ten-platform [`runtime-llamacpp-v0.2.6`](https://github.com/modelscope/FunASR/releases/tag/runtime-llamacpp-v0.2.6) archives cover Linux, macOS, and Windows CPU/GPU variants, including a dedicated Windows CUDA architecture 120 package for RTX 50 / Blackwell GPUs.- **Joint transcription and diarization** — the third-party [MOSS-Transcribe-Diarize](./moss_transcribe_diarize.md) model is available through `AutoModel` with local Transformers and vLLM backends, or as an independent native SGLang Omni service. SGLang Omni is not an `AutoModel` backend. The model produces timestamps and speaker labels in one pass without a separate external VAD or speaker model; OpenMOSS remains the model owner.-- **Repository roles and issue routing** — [#3203](https://github.com/modelscope/FunASR/issues/3203) tracks this document and the remaining model-weight and vLLM entry-point questions. It stays open until those questions have evidence and the reporter has time to confirm.+- **Repository roles and issue routing** - [#3203](https://github.com/modelscope/FunASR/issues/3203) tracks this document and contribution-entry feedback. The reporter has confirmed the vLLM A/B entry-point explanation; the broader roadmap issue remains open, and documentation delivery does not close it automatically.### In progress-- **Fun-ASR-Nano native Transformers integration** — [huggingface/transformers#46180](https://github.com/huggingface/transformers/pull/46180) is in review; use the PR's exact-head CI and review state as the source of truth.+- **Native Transformers downstream compatibility** - keep the [native model entry point](https://github.com/QwenAudio/Fun-ASR) and downstream transcription regressions aligned with exact model and library revisions. The initial integration is delivered; compatibility work is not a pending initial merge.- **Restore complete public checkpoint functionality** — [#3496](https://github.com/modelscope/FunASR/issues/3496) tracks missing CTC tensors needed by timestamp and diarization paths in the Hugging Face checkpoint.- **Realtime preview efficiency and L20 validation** — [#3528](https://github.com/modelscope/FunASR/issues/3528) established that v1.3.9 appeared faster by silently skipping most partial previews while its event loop was blocked. The issue remains open for equal-work L20 profiling and a deliberate refresh/window policy; it is not treated as a resolved throughput regression.-- **Qwen3-ASR offline vLLM workflow** — [#3592](https://github.com/modelscope/FunASR/pull/3592) adds a tested native `Qwen3ASRModel.LLM` example. [#3419](https://github.com/modelscope/FunASR/issues/3419) remains open until the reporter's 8–9% CER result can be reproduced with an exact model revision, service configuration, and scoring script.-- **AMD Windows Vulkan validation** — [#3479](https://github.com/modelscope/FunASR/issues/3479) remains open for reporter hardware retesting against `runtime-llamacpp-v0.2.6`; publication of the archive is not evidence that the hardware crash is fixed.+- **AMD Windows Vulkan validation** - [#3479](https://github.com/modelscope/FunASR/issues/3479) remains open for the original affected hardware paths. A reporter verified successful RX 9070 XT execution with runtime v0.2.5 and AMD 26.8.1 together; the driver also changed, so this does not isolate the runtime fix or establish the original 7600M XT / 780M outcome. Published archives alone are not hardware acceptance evidence.### Nextdiff --git a/docs/repository_roles_zh.md b/docs/repository_roles_zh.mdindex 531d7a703..0625b21a9 100644--- a/docs/repository_roles_zh.md+++ b/docs/repository_roles_zh.md@@ -91,7 +91,7 @@| [L20 等 GPU 上的实时预览效率](https://github.com/modelscope/FunASR/issues/3528) | 对齐 partial 消息数量后,怎样选择刷新间隔和 partial window,才能在不静默跳过预览的前提下取得合适的延迟/吞吐平衡? | 基于 exact commit 的客户端 JSONL 和 `--log-decode-profile` 服务端日志;固定 SPK、ping、音频、并发数、partial window 与 partial 消息数量 | 在 L20、L4、A10 或其他非 H100 GPU 上复现,并分析 queue、encoder 与 engine 时间 || [AMD Windows Vulkan 稳定性](https://github.com/modelscope/FunASR/issues/3479) | 当前 runtime 能否在报告者的 AMD GPU 上完成模型初始化和转写;若不能,最后成功的初始化边界在哪里? | 精确压缩包名称和 SHA256、GPU/驱动/Windows 版本、完整初始化日志,以及报告者硬件复测 | AMD Windows 硬件所有者和 Vulkan/llama.cpp 贡献者 || [恢复公开 checkpoint 的完整能力](https://github.com/modelscope/FunASR/issues/3496) | 如何从有权限的模型所有者账号发布缺失 CTC tensors,并在上传后完成验证? | 不可变模型 revision、文件哈希、公开 clean-cache 回下载和真实时间戳/说话人推理 | 有 Hugging Face 写权限的模型所有者和 checkpoint 验证贡献者 |-| [上游模型集成](https://github.com/huggingface/transformers/pull/46180) | 如何让 Fun-ASR-Nano 保持 Transformers 上游兼容,同时保留固定的 model card 和回归测试边界? | exact-head 上游 CI、聚焦本地测试、model card review 与维护者 review | Transformers reviewer,以及能在合并前验证下游加载的用户 |+| [原生 Transformers 兼容性](https://github.com/QwenAudio/Fun-ASR) | 已合并的 Fun-ASR-Nano 集成能否在支持的 Transformers 版本中保持下游加载和转写兼容? | 精确模型与库版本、聚焦转写测试,以及 model card 的能力边界 | 能复现下游兼容性问题的 Transformers 用户 |### 认领前@@ -110,19 +110,20 @@### 已交付+- **Fun-ASR-Nano 的 Transformers 原生集成** - [huggingface/transformers#46180](https://github.com/huggingface/transformers/pull/46180) 已于 2026-09-09 合并。请使用[原生 `-hf` checkpoint](https://huggingface.co/FunAudioLLM/Fun-ASR-Nano-2512-hf)与[模型仓库的 Transformers 入口](https://github.com/QwenAudio/Fun-ASR)。原生生成路径不包含工具包的辅助 CTC 分支;这项交付不代表已有原生时间戳、说话人区分或实时会话支持。+- **Qwen3-ASR 离线 vLLM 工作流** - [#3592](https://github.com/modelscope/FunASR/pull/3592) 已交付原生 `Qwen3ASRModel.LLM` 示例。报告者确认完成,并于 2026-09-09 关闭 [#3419](https://github.com/modelscope/FunASR/issues/3419)。其中的准确率对照基于该报告的音频、参考文本与自定义评分脚本,不是通用 CER 保证。- **实时服务长会话状态有界** —— [#3214](https://github.com/modelscope/FunASR/pull/3214) 与 [QwenAudio/Fun-ASR#135](https://github.com/QwenAudio/Fun-ASR/pull/135) 已合并,诊断能力已发布,报告者证据使 [#3101](https://github.com/modelscope/FunASR/issues/3101) 可以关闭。- **稳定的应用接口** —— 工具包现已提供 OpenAI-compatible 转写服务、健康检查、浏览器与命令行 smoke test,并在[部署矩阵](./deployment_matrix_zh.md)中列出 Python / CLI / HTTP / WebSocket 入口。- **工业与边缘部署路径** —— vLLM 服务和签名发布流程已有文档;经验证的十平台 [`runtime-llamacpp-v0.2.6`](https://github.com/modelscope/FunASR/releases/tag/runtime-llamacpp-v0.2.6) 压缩包覆盖 Linux、macOS 与 Windows 的 CPU/GPU 变体,并提供面向 RTX 50 / Blackwell 的 Windows CUDA architecture 120 专用包。- **联合转写与说话人识别** —— 第三方 [MOSS-Transcribe-Diarize](./moss_transcribe_diarize_zh.md) 模型已通过 `AutoModel` 接入本地 Transformers 与 vLLM 后端,也可通过原生 SGLang Omni 独立服务。SGLang Omni 不是 `AutoModel` backend。它在一次推理中生成时间戳与说话人标签,不需要额外的外部 VAD 或说话人模型;模型所有者仍是 OpenMOSS。-- **仓库职责与 issue 路由** —— [#3203](https://github.com/modelscope/FunASR/issues/3203) 继续跟踪本文档以及尚未回答完的模型权重和 vLLM 入口问题。在这些问题有证据且报告者有合理确认时间之前,issue 保持开放。+- **仓库职责与 issue 路由** - [#3203](https://github.com/modelscope/FunASR/issues/3203) 跟踪本文档与贡献入口反馈。报告者已确认 vLLM A/B 入口说明清楚;更广泛的路线图 issue 仍保持开放,不因文档交付而自动关闭。### 进行中-- **Fun-ASR-Nano 的 Transformers 原生集成** —— [huggingface/transformers#46180](https://github.com/huggingface/transformers/pull/46180) 正在审查;以该 PR 的 exact-head CI 与 review 状态为准。+- **原生 Transformers 下游兼容性** - 按精确模型与库版本持续核对[原生模型入口](https://github.com/QwenAudio/Fun-ASR)与下游转写回归。初始集成已经交付;兼容性维护不代表首次合并仍在等待。- **恢复公开 checkpoint 的完整能力** —— [#3496](https://github.com/modelscope/FunASR/issues/3496) 跟踪 Hugging Face checkpoint 缺少时间戳与说话人路径所需 CTC tensors 的问题。- **实时预览效率与 L20 验证** —— [#3528](https://github.com/modelscope/FunASR/issues/3528) 已确认 v1.3.9 看似更快,是因为事件循环阻塞时静默跳过了大部分 partial 预览。该 issue 继续开放,用于等工作量的 L20 profiling 和明确的刷新/window 策略;不能把它当成已经解决的吞吐回退。-- **Qwen3-ASR 离线 vLLM 工作流** —— [#3592](https://github.com/modelscope/FunASR/pull/3592) 增加经过验证的原生 `Qwen3ASRModel.LLM` 示例。[#3419](https://github.com/modelscope/FunASR/issues/3419) 继续开放,直到能用精确模型 revision、服务配置和评分脚本复现报告者的 8–9% CER。-- **AMD Windows Vulkan 验证** —— [#3479](https://github.com/modelscope/FunASR/issues/3479) 保持开放,等待报告者在 `runtime-llamacpp-v0.2.6` 上进行硬件复测;发布压缩包不等于硬件崩溃已经修复。+- **AMD Windows Vulkan 验证** - [#3479](https://github.com/modelscope/FunASR/issues/3479) 保持开放,等待最初受影响的硬件路径确认。已有报告者在 runtime v0.2.5 与 AMD 26.8.1 的组合上验证 RX 9070 XT 成功运行;驱动也同时改变,因此不能单独归因于 runtime 修复,更不能据此确认原始 7600M XT / 780M 的结果。发布压缩包本身不等于硬件验收。### 下一步diff --git a/tests/test_repository_roadmap_docs.py b/tests/test_repository_roadmap_docs.pynew file mode 100644index 000000000..c983d5f2e--- /dev/null+++ b/tests/test_repository_roadmap_docs.py@@ -0,0 +1,46 @@+"""Keep completed roadmap milestones separate from unresolved validation."""++from pathlib import Path+import unittest+++ROOT = Path(__file__).resolve().parents[1]+GUIDES = (+ ("repository_roles.md", "### Delivered", "### In progress", "### Next"),+ ("repository_roles_zh.md", "### \u5df2\u4ea4\u4ed8", "### \u8fdb\u884c\u4e2d", "### \u4e0b\u4e00\u6b65"),+)+++class RepositoryRoadmapContract(unittest.TestCase):+ def sections(self, guide):+ name, delivered, active, following = guide+ text = (ROOT / "docs" / name).read_text(encoding="utf-8")+ return (text.split(delivered, 1)[1].split(active, 1)[0],+ text.split(active, 1)[1].split(following, 1)[0])++ def test_native_transformers_merge_is_delivered(self):+ for guide in GUIDES:+ with self.subTest(guide=guide[0]):+ delivered, active = self.sections(guide)+ self.assertIn("huggingface/transformers/pull/46180", delivered)+ self.assertIn("FunAudioLLM/Fun-ASR-Nano-2512-hf", delivered)+ self.assertNotIn("huggingface/transformers/pull/46180", active)++ def test_reporter_closed_qwen_workflow_is_delivered(self):+ for guide in GUIDES:+ with self.subTest(guide=guide[0]):+ delivered, active = self.sections(guide)+ self.assertIn("FunASR/pull/3592", delivered)+ self.assertIn("FunASR/issues/3419", delivered)+ self.assertNotIn("FunASR/issues/3419", active)++ def test_unresolved_hardware_and_checkpoint_work_stays_open(self):+ for guide in GUIDES:+ with self.subTest(guide=guide[0]):+ _, active = self.sections(guide)+ for issue in (3496, 3528, 3479):+ self.assertIn(f"FunASR/issues/{issue}", active)+++if __name__ == "__main__":+ unittest.main()