Add transcription model selection and lifecycle management

README.md

@@ -320,14 +320,20 @@ precedence on NVIDIA systems.
 
 `large-v3-turbo` is the modern speed/quality sweet spot at the top end (≈8× faster
 decoding than `large-v3`); pick it if accuracy matters more than leaving the CPU idle.
-Drop the file in `resources/whisper/models/` and set its filename in Settings, or pass
-it to the script: `pwsh ./scripts/setup-whisper.ps1 -Model ggml-large-v3-turbo-q5_0.bin`.
+Drop the file in `resources/whisper/models/`; Parley discovers installed `.bin` files
+and lists each one under **Settings → Transcription → Model**. You can also pass it to
+the setup script: `pwsh ./scripts/setup-whisper.ps1 -Model ggml-large-v3-turbo-q5_0.bin`.
 
 ### Or: use a remote transcription server
 
-If you'd rather not transcribe on this machine, run a compatible server elsewhere
-and set **Settings → Transcription → Remote transcription URL** (e.g.
-`http://192.168.1.10:8765`). When set, Parley skips the bundled engine entirely.
+If you'd rather not transcribe on this machine, run a compatible server elsewhere,
+select **Settings → Transcription → Model → External server**, and enter its URL
+(e.g. `http://192.168.1.10:8765`). Parley can test reachability but does not manage
+the remote process. When selected, it skips local model loading entirely.
+
+The same section lists Automatic, installed Nemotron, and every installed Whisper
+model. Local selections can be started, stopped to release memory, or restarted while
+no meeting is active. Starting a meeting automatically reloads a stopped local model.
 
 > ⚠️ **Platform note:** the bundled-engine path is currently hard-coded to the Windows
 > layout (`bin/Release/whisper-server.exe`). On macOS/Linux, use the **remote URL**
@@ -369,8 +375,8 @@ Nemotron model or Python/CUDA runtime.
 3. **Settings** (gear icon): save one **LLM connection** per provider (name, base
    URL, model, optional API key) — a local llama-server / LM Studio / Ollama, or a
    cloud URL. Mark one **active** (★), **Test** each, and set the analysis interval
-   and transcription options. Switch which connection a meeting uses from the
-   **LLM connection dropdown in the header** (before you start the meeting).
+   and choose/manage the transcription model. Switch which LLM connection a meeting
+   uses from the **LLM connection dropdown in the header** (before you start the meeting).
 4. Check the footer for the installed version and selected **Voice-to-text** model.
    Local model weights begin loading when Parley opens; if the footer still says
    *Loading local model…*, starting a meeting waits only for the remaining load time.
@@ -414,23 +420,23 @@ pre-meeting context followed by every timestamped transcript line.
 
 - **"The local transcription engine isn't installed" on Start.** You haven't fetched
   the whisper engine yet — run **`task setup:whisper`** (or `scripts/setup-whisper.ps1`),
-  or set a remote transcription URL in Settings. Parley shows the reason in a red banner
+  or select an External server in Settings. Parley shows the reason in a red banner
   and writes full details to **`parley.log`** in your app-data folder (Windows:
   `%AppData%\Parley\`). For a packaged build, the `resources/whisper/` folder must sit
   next to the `.exe`.
 - **The installer says no NVIDIA GPU, but `nvidia-smi -L` shows one.** Install
   Parley **v0.1.3 or newer**. Older installers ran the 64-bit NVIDIA utility through
   a redirected 32-bit shell, which could incorrectly report no GPU.
-- **Nemotron was not selected on an NVIDIA system.** The footer shows the backend
-  Parley actually selected. Check
+- **Automatic did not select Nemotron on an NVIDIA system.** The footer shows the
+  backend Parley actually selected. Check
   `%AppData%\Parley\nemotron-server.log`. Parley requires a complete
-  `resources\nemotron` installation with a `.ready` marker and falls back to CPU
-  Whisper when the model cannot load. On an installed per-user copy, close Parley
-  and resume provisioning from 64-bit PowerShell (the script reuses files already
-  present):
+  Nemotron installation with a `.ready` marker and falls back to CPU Whisper when
+  the model cannot load. An explicitly selected Nemotron reports the failure instead
+  of silently switching models. On an installed per-user copy, close Parley and
+  resume provisioning from 64-bit PowerShell; the script reuses files already present:
 
   ```powershell
-  powershell.exe -NoProfile -ExecutionPolicy Bypass -File "$env:LOCALAPPDATA\Programs\Parley\resources\nemotron\setup.ps1" -InstallRoot "$env:LOCALAPPDATA\Programs\Parley\resources\nemotron"
+  powershell.exe -NoProfile -ExecutionPolicy Bypass -File "$env:LOCALAPPDATA\Programs\Parley\resources\nemotron\setup.ps1" -InstallRoot "$env:LOCALAPPDATA\Parley\nemotron"
   ```
 - **"No mic" with a mic selected.** The badge now reflects whether a microphone source
   actually started. If it still says *No mic*, that device failed to open (wrong device,

frontend/bindings/github.com/tomvokac/parley/index.ts

@@ -12,5 +12,7 @@ export type {
     AnalysisStatusEvent,
     LoadedSession,
     RuntimeInfo,
-    StatusEvent
+    StatusEvent,
+    TranscriptionConfig,
+    TranscriptionModelOption
 } from "./models.js";

frontend/bindings/github.com/tomvokac/parley/internal/store/models.ts

@@ -108,9 +108,15 @@ export interface Settings {
     "hasAPIKey": boolean;
     "captureSources": CaptureSource[] | null;
 
+    /**
+     * SttEngine selects how transcription is provided: "auto", "nemotron",
+     * "whisper", or "external". Auto preserves the original GPU-first fallback.
+     */
+    "sttEngine": string;
+
     /**
      * SttBaseURL, when set, points transcription at a remote /inference-compatible
-     * server (e.g. http://host:8765) instead of launching a local engine.
+     * server (e.g. http://host:8765). It is used only when SttEngine is "external".
      */
     "sttBaseURL": string;
 

frontend/bindings/github.com/tomvokac/parley/meetingservice.ts

@@ -30,6 +30,14 @@ export function AddLiveNote(scope: string, text: string): $CancellablePromise<st
     return $Call.ByID(1344950533, scope, text);
 }
 
+/**
+ * ConfigureTranscription validates and persists a model selection, then
+ * reconciles the idle runtime. Local loading continues asynchronously.
+ */
+export function ConfigureTranscription(config: $models.TranscriptionConfig): $CancellablePromise<void> {
+    return $Call.ByID(713856948, config);
+}
+
 /**
  * DeleteSession permanently removes a saved meeting (not the one in progress).
  */
@@ -85,6 +93,14 @@ export function ListSessions(): $CancellablePromise<store$0.Session[] | null> {
     return $Call.ByID(1774945289);
 }
 
+/**
+ * ListTranscriptionModels discovers the selectable local installations and the
+ * always-available external-server choice.
+ */
+export function ListTranscriptionModels(): $CancellablePromise<$models.TranscriptionModelOption[] | null> {
+    return $Call.ByID(1026157818);
+}
+
 /**
  * LoadSession returns a saved meeting's full state for display.
  */
@@ -99,6 +115,13 @@ export function RenameSession(id: number, title: string): $CancellablePromise<vo
     return $Call.ByID(3925232594, id, title);
 }
 
+/**
+ * RestartTranscriptionModel replaces the current local process with a fresh load.
+ */
+export function RestartTranscriptionModel(): $CancellablePromise<void> {
+    return $Call.ByID(3040477430);
+}
+
 /**
  * Resume continues a previously saved meeting, appending new transcript/analysis
  * to it. Pass the session id returned by ListSessions.
@@ -115,9 +138,32 @@ export function Start(): $CancellablePromise<void> {
     return $Call.ByID(4124033072);
 }
 
+/**
+ * StartTranscriptionModel loads the selected local model without starting a meeting.
+ */
+export function StartTranscriptionModel(): $CancellablePromise<void> {
+    return $Call.ByID(2331669113);
+}
+
 /**
  * Stop ends the session, flushing the final audio and closing recordings.
  */
 export function Stop(): $CancellablePromise<void> {
     return $Call.ByID(252224180);
 }
+
+/**
+ * StopTranscriptionModel releases the selected local model while Parley is idle.
+ */
+export function StopTranscriptionModel(): $CancellablePromise<void> {
+    return $Call.ByID(207813981);
+}
+
+/**
+ * TestExternalTranscription verifies that an external HTTP server is reachable.
+ * Any HTTP response counts as reachable; the inference path is exercised by the
+ * first real transcription request.
+ */
+export function TestExternalTranscription(baseURL: string): $CancellablePromise<void> {
+    return $Call.ByID(1170675977, baseURL);
+}

frontend/bindings/github.com/tomvokac/parley/models.ts

@@ -42,11 +42,18 @@ export interface LoadedSession {
 export interface RuntimeInfo {
     "appVersion": string;
     "transcriptionModel": string;
+    "transcriptionModelID": string;
 
     /**
-     * loading | ready | error
+     * local | external
+     */
+    "transcriptionKind": string;
+
+    /**
+     * stopped | loading | ready | configured | error
      */
     "transcriptionStatus": string;
+    "transcriptionMessage": string;
 }
 
 /**
@@ -61,3 +68,28 @@ export interface StatusEvent {
     "micAvailable": boolean;
     "activeSources": string[] | null;
 }
+
+/**
+ * TranscriptionConfig is the user-facing selection saved by the model manager.
+ * ModelID is one of auto, nemotron, whisper:<filename>, or external.
+ */
+export interface TranscriptionConfig {
+    "modelID": string;
+    "externalURL": string;
+}
+
+/**
+ * TranscriptionModelOption describes one selectable transcription provider.
+ */
+export interface TranscriptionModelOption {
+    "id": string;
+    "label": string;
+
+    /**
+     * automatic | local | external
+     */
+    "kind": string;
+    "detail": string;
+    "available": boolean;
+    "unavailableReason": string;
+}

frontend/src/App.tsx

@@ -3,6 +3,7 @@ import { Events } from "@wailsio/runtime";
 import {
   AlertTriangle,
   AudioLines,
+  CheckCircle2,
   Download,
   History,
   Loader2,
@@ -152,7 +153,10 @@ function App() {
   const [runtimeInfo, setRuntimeInfo] = useState<RuntimeInfo>({
     appVersion: "",
     transcriptionModel: "Loading local model…",
+    transcriptionModelID: "auto",
+    transcriptionKind: "local",
     transcriptionStatus: "loading",
+    transcriptionMessage: "Loading the selected transcription model…",
   });
   // Session-scoped suggestion pins/dismissals, keyed by normalized text. The
   // `analysis` listener is registered once, so it reads these via refs to avoid a
@@ -265,6 +269,7 @@ function App() {
           ...info,
           transcriptionModel: "Model status unavailable",
           transcriptionStatus: "error",
+          transcriptionMessage: "Parley could not read the transcription model status.",
         }))
       );
 
@@ -708,11 +713,20 @@ function App() {
         </span>
         <span
           className="flex min-w-0 items-center gap-1.5"
-          title={`Voice-to-text model: ${runtimeInfo.transcriptionModel}`}
+          title={`Voice-to-text model: ${runtimeInfo.transcriptionModel}${runtimeInfo.transcriptionMessage ? ` — ${runtimeInfo.transcriptionMessage}` : ""}`}
         >
           {runtimeInfo.transcriptionStatus === "loading" && (
             <Loader2 className="h-3 w-3 shrink-0 animate-spin" />
           )}
+          {(runtimeInfo.transcriptionStatus === "ready" || runtimeInfo.transcriptionStatus === "configured") && (
+            <CheckCircle2 className="h-3 w-3 shrink-0 text-emerald-400" />
+          )}
+          {runtimeInfo.transcriptionStatus === "error" && (
+            <AlertTriangle className="h-3 w-3 shrink-0 text-destructive" />
+          )}
+          {runtimeInfo.transcriptionStatus === "stopped" && (
+            <Square className="h-2.5 w-2.5 shrink-0" />
+          )}
           <span className="shrink-0">Voice-to-text:</span>
           <span className="truncate text-foreground/75">
             {runtimeInfo.transcriptionModel}
@@ -721,7 +735,12 @@ function App() {
       </footer>
 
       <AudioDialog open={audioOpen} onOpenChange={setAudioOpen} />
-      <SettingsDialog open={settingsOpen} onOpenChange={setSettingsOpen} />
+      <SettingsDialog
+        open={settingsOpen}
+        onOpenChange={setSettingsOpen}
+        meetingActive={active}
+        runtimeInfo={runtimeInfo}
+      />
       <ContextDialog open={contextOpen} onOpenChange={setContextOpen} />
       <SessionsDialog
         open={sessionsOpen}