Pr/sherpa streaming backend

[weikeyi]

User description

摘要

• 将 Windows sherpa-onnx-local 从骨架推进到实际 runtime/provider 实现，支持 offline batch 与 experimental online streaming 两种模型模式。
• 新增 Zipformer 中英双语流式模型 catalog、必需文件、下载映射和 mode 判断。
• 后端 runtime 分离缓存 OfflineRecognizer 与 OnlineRecognizer，并记录最近一次 prepare/transcribe/audio/error 诊断信息。
• provider 增加 online worker：录音时消费 PCM chunk，输出 partial token，停止录音后返回最终 RawTranscript。
• Coordinator/commands 接入 sherpa streaming：online partial 通过 local-asr-token 事件发出，final 继续复用现有 polish / insert / history 收尾路径。
• 增加 sherpa 模型 alias、language hint、下载校验、release archive、runtime 状态等单元测试覆盖。

验证

• cargo check --manifest-path src-tauri/Cargo.toml
• cargo test --manifest-path src-tauri/Cargo.toml --lib sherpa

结果：sherpa 相关测试 42 passed; 0 failed。

说明

• 这个 PR 只包含后端 runtime/provider/Coordinator/commands 接线，不包含前端 UI、i18n、spike example 和规划文档。
• online streaming 代码路径已接入，但仍需要 Windows 真机下载模型后验证 partial/final、RTF、CPU 和长录音稳定性。

---

PR Type

Enhancement, Tests

---

Description

• Add Zipformer streaming model catalog

• Implement offline and online recognizers

• Emit partial tokens during dictation

• Harden downloads, archives, and status tests

---

Diagram Walkthrough

flowchart LR
  A["Sherpa catalog and metadata"] --> B["Sherpa runtime"]
  C["Download and archive checks"] --> B
  B --> D["Sherpa provider"]
  D --> E["Coordinator dictation"]
  D -- "partial tokens" --> F["local-asr-token events"]
  E -- "final transcript" --> G["History / polish flow"]

Loading

File Walkthrough

	Relevant files
Documentation	2 files mod.rs Document sherpa offline and streaming modes                            +1/-1      types.rs Refresh sherpa preference documentation                                    +1/-2
Enhancement	5 files sherpa.rs Add streaming model metadata and helpers                                  +88/-9    sherpa_provider.rs Split provider into offline and online workers                      +273/-20 sherpa_runtime.rs Implement offline and streaming recognizer lifecycle          +590/-51 coordinator.rs Update coordinator for sherpa runtime lifecycle                    +7/-9      dictation.rs Wire streaming partial tokens into dictation                          +35/-8
Tests	2 files sherpa_download.rs Verify archives and download integrity                                      +186/-0  commands.rs Add sherpa alias and hint validation tests                              +53/-10

---

[github-actions]

PR Reviewer Guide 🔍

Here are some key observations to aid the review process:

⏱️ Estimated effort to review: 3 🔵🔵🔵⚪⚪

🧪 PR contains tests

🔒 No security concerns identified

⚡ Recommended focus areas for review Timeout Leak If the online recognizer exceeds audio_timeout, finish() returns an error after sending Cancel but drops the worker JoinHandle without joining it. A slow or stuck streaming session can therefore leave the background thread running until process exit, leaking resources across repeated timeouts. async fn finish(self, audio_timeout: Duration) -> Result<String> { let _ = self.tx.send(OnlineWorkerMessage::Finish); let result_rx = self .result_rx .lock() .take() .ok_or_else(|| anyhow::anyhow!("sherpa-onnx streaming result already taken"))?; let join_handle = self.join_handle.lock().take(); let result = tokio::time::timeout(audio_timeout, async move { tokio::task::spawn_blocking(move || { result_rx.recv().map_err(|error| { anyhow::anyhow!("sherpa-onnx streaming worker closed: {error}") })? }) .await .map_err(|error| anyhow::anyhow!("sherpa-onnx streaming join failed: {error:#}"))? }) .await; let result = match result { Ok(result) => result, Err(_) => { self.cancelled.store(true, Ordering::SeqCst); let _ = self.tx.send(OnlineWorkerMessage::Cancel); anyhow::bail!("sherpa-onnx streaming transcribe timeout"); } }; if let Some(join_handle) = join_handle { let _ = join_handle.join(); } Unbounded Queue The online path uses an unbounded mpsc::channel for PCM chunks. If the recognizer falls behind the recorder on a slow CPU or during a long session, audio buffers can accumulate without limit and grow memory usage linearly with recording length. let (tx, rx) = mpsc::channel::<OnlineWorkerMessage>(); let (result_tx, result_rx) = mpsc::channel::<Result<String>>();

[H-Chris233]

如果实测没有问题的话，踢我一下我给你合了

[weikeyi]

@H-Chris233 实测过 流式输出没有问题