Add Gemini 3.5 Transcribe as a BYOK speech-to-text provider

apps/desktop/src/settings/ai/stt/selection.test.ts

@@ -18,6 +18,9 @@ describe("getDefaultSttModel", () => {
     expect(getDefaultSttModel("groq")).toBe("whisper-large-v3-turbo");
     expect(getDefaultSttModel("openrouter")).toBe("openai/gpt-transcribe");
     expect(getDefaultSttModel("xai")).toBe("xai-stt");
+    expect(getDefaultSttModel("google_generative_ai")).toBe(
+      "gemini-3.5-transcribe-live",
+    );
     expect(getDefaultSttModel("together")).toBe("openai/whisper-large-v3");
     expect(getDefaultSttModel("speechmatics")).toBe("enhanced");
     expect(getDefaultSttModel("azure_speech")).toBe("fast-transcription");

apps/desktop/src/settings/ai/stt/shared.tsx

@@ -8,8 +8,8 @@ import {
   Cohere,
   ElevenLabs,
   Fireworks,
-  GoogleCloud,
   Gemini,
+  GoogleCloud,
   Groq,
   Mistral,
   OpenAI,

crates/listener2-core/src/lib.rs

@@ -172,7 +172,7 @@ pub fn suggest_providers_for_languages_batch(languages: &[anlg_language::Languag
 
     with_support
         .into_iter()
-        .map(|(kind, _)| format!("{:?}", kind).to_lowercase())
+        .map(|(kind, _)| kind.to_string())
         .collect()
 }
 

crates/owhisper-client/src/adapter/google_generative_ai/live.rs

@@ -254,12 +254,18 @@ fn transcription_response(
         })
         .collect();
 
+    let from_finalize = is_final
+        && value
+            .get("finished")
+            .and_then(serde_json::Value::as_bool)
+            .unwrap_or(false);
+
     Some(StreamResponse::TranscriptResponse {
         start,
         duration: (end - start).max(0.0),
         is_final,
         speech_final: is_final,
-        from_finalize: false,
+        from_finalize,
         channel: Channel {
             alternatives: vec![Alternatives {
                 transcript: text.to_string(),
@@ -405,6 +411,14 @@ mod tests {
         assert_eq!(channel.alternatives[0].words[0].speaker, Some(2));
         assert_eq!(*start, 0.1);
         assert!((*duration - 0.7).abs() < f64::EPSILON);
+
+        let finalize_flush = adapter.parse_response(
+            r#"{"serverContent":{"inputTranscription":{"text":"hello there","finished":true}}}"#,
+        );
+        let StreamResponse::TranscriptResponse { from_finalize, .. } = &finalize_flush[0] else {
+            panic!("expected finalize flush");
+        };
+        assert!(*from_finalize);
     }
 
     #[test]

crates/owhisper-client/src/providers.rs

@@ -326,11 +326,11 @@ impl Provider {
             Self::AwsTranscribe
             | Self::AzureSpeech
             | Self::GoogleCloud
+            | Self::GoogleGenerativeAi
             | Self::Groq
             | Self::RevAi
             | Self::Speechmatics
             | Self::Together
-            | Self::GoogleGenerativeAi
             | Self::Xai => None,
         }
     }

packages/changelog/content/1.4.14.md

@@ -23,6 +23,8 @@ summary: "Anarlog 1.4.14 keeps live meetings going through brief network drops,
 
 - Choose a local whisper.cpp `.bin` model file from Settings → Transcription
   instead of relying only on bundled downloads
+- Transcribe live with Google Gemini 3.5 Transcribe, then refine the recording
+  with speaker labels and word timestamps using the same API key
 - Keep Mistral batch transcription speaker labels aligned with the words they
   belong to
 - Load ChatGPT subscription models from the current catalog instead of a

plugins/transcription/src/listener/commands.rs

@@ -164,7 +164,7 @@ pub async fn suggest_providers_for_languages_live<R: tauri::Runtime>(
 
     let supported: Vec<String> = with_support
         .into_iter()
-        .map(|(kind, _)| format!("{:?}", kind).to_lowercase())
+        .map(|(kind, _)| kind.to_string())
         .collect();
 
     Ok(supported)