@mastra/voice-google-gemini-live 0.11.5-alpha.2 → 0.12.0

package/CHANGELOG.md

@@ -1,5 +1,108 @@
 # @mastra/voice-google-gemini-live
 
+## 0.12.0
+
+### Minor Changes
+
+- Surface native-audio behavioral signals on Gemini Live realtime sessions (#17021). ([#17434](https://github.com/mastra-ai/mastra/pull/17434))
+
+  The `@mastra/voice-google-gemini-live` provider now enables transcription and barge-in detection in the setup payload and exposes them through Mastra's standard realtime event contract. This makes native-audio models such as `gemini-2.5-flash-native-audio-preview-12-2025` and `gemini-3.1-flash-live-preview` behaviorally usable end-to-end. Until now, the spoken response was silently dropped on native-audio because it arrives on a different wire channel from the model's internal reasoning.
+
+  **What changed**
+  - Setup payload unconditionally includes `input_audio_transcription`, `output_audio_transcription`, and `realtime_input_config.activity_handling = 'START_OF_ACTIVITY_INTERRUPTS'`, matching how the OpenAI, xAI, Inworld, and AWS Nova Sonic providers enable transcription by default.
+  - User-side transcripts emit as `writing` with `role: 'user'`. Model-side transcripts emit as `writing` with `role: 'assistant'`. This matches the cross-provider `writing` contract.
+  - Barge-in (the server cancelling its in-flight response when the user starts speaking) emits an `interrupt` event with `{ type: 'user', timestamp }`, matching `@mastra/voice-aws-nova-sonic`.
+  - On native-audio models, `modelTurn.parts.text` is the model's internal chain-of-thought, not the spoken response. It now emits as a Gemini-specific `thinking` event so consumers can render reasoning separately. On non-native-audio models, `modelTurn.parts.text` continues to emit as `writing` (it is the spoken response there).
+
+  **Example**
+
+  ```ts
+  import { GeminiLiveVoice } from '@mastra/voice-google-gemini-live';
+
+  const voice = new GeminiLiveVoice({
+    apiKey: process.env.GOOGLE_API_KEY,
+    model: 'gemini-2.5-flash-native-audio-preview-12-2025',
+  });
+
+  voice.on('writing', ({ text, role }) => {
+    // role: 'user'      → speech-to-text of the caller
+    // role: 'assistant' → speech-to-text of the model's spoken reply
+  });
+
+  voice.on('thinking', ({ text }) => {
+    // Gemini's internal reasoning on native-audio models
+  });
+
+  voice.on('interrupt', ({ type, timestamp }) => {
+    // Drop queued TTS audio — the user just barged in
+  });
+
+  await voice.connect();
+  ```
+
+### Patch Changes
+
+- Moved shared voice primitives and route metadata into the new `@internal/voice` package so voice providers no longer depend on `@mastra/core` and server voice routes share the same route definitions. ([#16725](https://github.com/mastra-ai/mastra/pull/16725))
+
+  `@mastra/core/voice` continues to re-export the voice APIs for backwards compatibility.
+
+- Fixed Gemini Live tool registration failing with `1007 Unknown name` errors for tools using discriminated unions, literals, and nullable types. The `sanitizeToolParameters` method now rewrites `oneOf` → `anyOf`, `const` → `enum`, and collapses nullable `anyOf` patterns into OpenAPI 3.0-compatible `type` + `nullable: true` form. Fixes #17020. ([#17179](https://github.com/mastra-ai/mastra/pull/17179))
+
+- **Fixed** Gemini Live sessions now connect successfully when using native-audio models. Previously the connection failed during session setup. ([#17019](https://github.com/mastra-ai/mastra/pull/17019))
+
+  **Fixed** tools are now invoked correctly. Previously tool calls were silently ignored even when tools were registered during setup.
+
+  **Fixed** tool results of any shape (arrays, primitives, objects) are now accepted. Previously, non-object tool return values caused sessions to close unexpectedly.
+
+  **Fixed** the `speaker` option is now honored when passed at the `VoiceConfig` root alongside `realtimeConfig`, not only when passed in the flat config shape.
+
+  **Changed** default model from `gemini-2.0-flash-exp` (shut down 2025-12-09) to `gemini-3.1-flash-live-preview` (Google's current Live API quickstart model). If you weren't explicitly setting `model`, your sessions will start connecting again.
+
+  Fixes #17018.
+
+- Updated dependencies [[`00eca42`](https://github.com/mastra-ai/mastra/commit/00eca4252393aa114dc8c9a5e1da68df91fa06cf), [`ff9d743`](https://github.com/mastra-ai/mastra/commit/ff9d743f71d7e072927725c0d700632aca0c1fee)]:
+  - @mastra/schema-compat@1.2.11
+
+## 0.12.0-alpha.3
+
+### Minor Changes
+
+- Surface native-audio behavioral signals on Gemini Live realtime sessions (#17021). ([#17434](https://github.com/mastra-ai/mastra/pull/17434))
+
+  The `@mastra/voice-google-gemini-live` provider now enables transcription and barge-in detection in the setup payload and exposes them through Mastra's standard realtime event contract. This makes native-audio models such as `gemini-2.5-flash-native-audio-preview-12-2025` and `gemini-3.1-flash-live-preview` behaviorally usable end-to-end. Until now, the spoken response was silently dropped on native-audio because it arrives on a different wire channel from the model's internal reasoning.
+
+  **What changed**
+  - Setup payload unconditionally includes `input_audio_transcription`, `output_audio_transcription`, and `realtime_input_config.activity_handling = 'START_OF_ACTIVITY_INTERRUPTS'`, matching how the OpenAI, xAI, Inworld, and AWS Nova Sonic providers enable transcription by default.
+  - User-side transcripts emit as `writing` with `role: 'user'`. Model-side transcripts emit as `writing` with `role: 'assistant'`. This matches the cross-provider `writing` contract.
+  - Barge-in (the server cancelling its in-flight response when the user starts speaking) emits an `interrupt` event with `{ type: 'user', timestamp }`, matching `@mastra/voice-aws-nova-sonic`.
+  - On native-audio models, `modelTurn.parts.text` is the model's internal chain-of-thought, not the spoken response. It now emits as a Gemini-specific `thinking` event so consumers can render reasoning separately. On non-native-audio models, `modelTurn.parts.text` continues to emit as `writing` (it is the spoken response there).
+
+  **Example**
+
+  ```ts
+  import { GeminiLiveVoice } from '@mastra/voice-google-gemini-live';
+
+  const voice = new GeminiLiveVoice({
+    apiKey: process.env.GOOGLE_API_KEY,
+    model: 'gemini-2.5-flash-native-audio-preview-12-2025',
+  });
+
+  voice.on('writing', ({ text, role }) => {
+    // role: 'user'      → speech-to-text of the caller
+    // role: 'assistant' → speech-to-text of the model's spoken reply
+  });
+
+  voice.on('thinking', ({ text }) => {
+    // Gemini's internal reasoning on native-audio models
+  });
+
+  voice.on('interrupt', ({ type, timestamp }) => {
+    // Drop queued TTS audio — the user just barged in
+  });
+
+  await voice.connect();
+  ```
+
 ## 0.11.5-alpha.2
 
 ### Patch Changes

package/README.md

@@ -99,10 +99,21 @@ voice.on('speaker', audioStream => {
 });
 
 voice.on('writing', ({ text, role }) => {
-  // Handle transcribed text
+  // role: 'user'      → speech-to-text of the caller
+  // role: 'assistant' → speech-to-text of the model's spoken reply
   console.log(`${role}: ${text}`);
 });
 
+// Native-audio models only: model's internal reasoning
+voice.on('thinking', ({ text }) => {
+  console.log(`thinking: ${text}`);
+});
+
+// Drop queued playback when the user barges in over the model
+voice.on('interrupt', ({ type, timestamp }) => {
+  console.log(`interrupt by ${type} at ${timestamp}`);
+});
+
 // Send text to speech
 await voice.speak('Hello from Mastra!');
 
@@ -313,16 +324,26 @@ Registers an event listener.
 
 - `'speaking'` - Audio response from model
 - `'speaker'` - Readable stream of concatenated audio for the active response
-- `'writing'` - Text response or transcription
+- `'writing'` - Transcribed text. Callback receives `{ text, role: 'user' | 'assistant' }`. On native-audio models the assistant transcript is driven by the server's `output_audio_transcription` channel
+- `'thinking'` - Model chain-of-thought / reasoning text on native-audio models. Callback receives `{ text }`. Does not fire on non-native-audio models, where reasoning is not surfaced separately
 - `'error'` - Error events
 - `'session'` - Session state changes
 - `'toolCall'` - Tool calls from model
 - `'vad'` - Voice activity detection events
-- `'interrupt'` - Interrupt events
+- `'interrupt'` - Emitted on barge-in when the user starts speaking over an in-flight model response. Callback receives `{ type: 'user', timestamp }`
 - `'usage'` - Token usage information
 - `'sessionHandle'` - Session resumption handle
 - `'turnComplete'` - Turn completion for the current model response
 
+#### Native-audio models
+
+Native-audio models (any model whose ID contains `native-audio`, e.g. `gemini-2.5-flash-native-audio-preview-12-2025`) split text output across two channels:
+
+- The model's spoken reply is delivered as audio plus an `output_audio_transcription` transcript — surfaced as `writing` with `role: 'assistant'`.
+- The model's internal reasoning is delivered as `modelTurn.parts.text` — surfaced as `thinking`.
+
+On non-native-audio models there is no `output_audio_transcription` channel; `modelTurn.parts.text` is the spoken response itself and is emitted as `writing` (so `thinking` will not fire). Transcription and barge-in detection are enabled automatically in the setup payload — no extra configuration is required.
+
 ### Tools
 
 Add tools with `addTools()` using either `@mastra/core/tools` or a plain object matching `ToolsInput`.

package/dist/docs/SKILL.md

@@ -3,7 +3,7 @@ name: mastra-voice-google-gemini-live
 description: Documentation for @mastra/voice-google-gemini-live. Use when working with @mastra/voice-google-gemini-live APIs, configuration, or implementation.
 metadata:
   package: "@mastra/voice-google-gemini-live"
-  version: "0.11.5-alpha.2"
+  version: "0.12.0"
 ---
 
 ## When to use

package/dist/docs/assets/SOURCE_MAP.json

@@ -1,5 +1,5 @@
 {
-  "version": "0.11.5-alpha.2",
+  "version": "0.12.0",
   "package": "@mastra/voice-google-gemini-live",
   "exports": {},
   "modules": {}

package/dist/docs/references/docs-voice-overview.md

@@ -16,7 +16,7 @@ const voiceAgent = new Agent({
   id: 'voice-agent',
   name: 'Voice Agent',
   instructions: 'You are a voice assistant that can help users with their tasks.',
-  model: 'openai/gpt-5.4',
+  model: 'openai/gpt-5.5',
   voice: new OpenAIVoice(),
 })
 ```
@@ -40,7 +40,7 @@ const voiceAgent = new Agent({
   id: 'voice-agent',
   name: 'Voice Agent',
   instructions: 'You are a voice assistant that can help users with their tasks.',
-  model: 'openai/gpt-5.4',
+  model: 'openai/gpt-5.5',
   voice: new OpenAIVoice(),
 })
 
@@ -68,7 +68,7 @@ const voiceAgent = new Agent({
   id: 'voice-agent',
   name: 'Voice Agent',
   instructions: 'You are a voice assistant that can help users with their tasks.',
-  model: 'openai/gpt-5.4',
+  model: 'openai/gpt-5.5',
   voice: new AzureVoice(),
 })
 
@@ -95,7 +95,7 @@ const voiceAgent = new Agent({
   id: 'voice-agent',
   name: 'Voice Agent',
   instructions: 'You are a voice assistant that can help users with their tasks.',
-  model: 'openai/gpt-5.4',
+  model: 'openai/gpt-5.5',
   voice: new ElevenLabsVoice(),
 })
 
@@ -122,7 +122,7 @@ const voiceAgent = new Agent({
   id: 'voice-agent',
   name: 'Voice Agent',
   instructions: 'You are a voice assistant that can help users with their tasks.',
-  model: 'openai/gpt-5.4',
+  model: 'openai/gpt-5.5',
   voice: new PlayAIVoice(),
 })
 
@@ -149,7 +149,7 @@ const voiceAgent = new Agent({
   id: 'voice-agent',
   name: 'Voice Agent',
   instructions: 'You are a voice assistant that can help users with their tasks.',
-  model: 'openai/gpt-5.4',
+  model: 'openai/gpt-5.5',
   voice: new GoogleVoice(),
 })
 
@@ -176,7 +176,7 @@ const voiceAgent = new Agent({
   id: 'voice-agent',
   name: 'Voice Agent',
   instructions: 'You are a voice assistant that can help users with their tasks.',
-  model: 'openai/gpt-5.4',
+  model: 'openai/gpt-5.5',
   voice: new CloudflareVoice(),
 })
 
@@ -203,7 +203,7 @@ const voiceAgent = new Agent({
   id: 'voice-agent',
   name: 'Voice Agent',
   instructions: 'You are a voice assistant that can help users with their tasks.',
-  model: 'openai/gpt-5.4',
+  model: 'openai/gpt-5.5',
   voice: new DeepgramVoice(),
 })
 
@@ -230,7 +230,7 @@ const voiceAgent = new Agent({
   id: 'voice-agent',
   name: 'Voice Agent',
   instructions: 'You are a voice assistant that can help users with their tasks.',
-  model: 'openai/gpt-5.4',
+  model: 'openai/gpt-5.5',
   voice: new InworldVoice(),
 })
 
@@ -257,7 +257,7 @@ const voiceAgent = new Agent({
   id: 'voice-agent',
   name: 'Voice Agent',
   instructions: 'You are a voice assistant that can help users with their tasks.',
-  model: 'openai/gpt-5.4',
+  model: 'openai/gpt-5.5',
   voice: new SpeechifyVoice(),
 })
 
@@ -284,7 +284,7 @@ const voiceAgent = new Agent({
   id: 'voice-agent',
   name: 'Voice Agent',
   instructions: 'You are a voice assistant that can help users with their tasks.',
-  model: 'openai/gpt-5.4',
+  model: 'openai/gpt-5.5',
   voice: new SarvamVoice(),
 })
 
@@ -311,7 +311,7 @@ const voiceAgent = new Agent({
   id: 'voice-agent',
   name: 'Voice Agent',
   instructions: 'You are a voice assistant that can help users with their tasks.',
-  model: 'openai/gpt-5.4',
+  model: 'openai/gpt-5.5',
   voice: new MurfVoice(),
 })
 
@@ -346,7 +346,7 @@ const voiceAgent = new Agent({
   id: 'voice-agent',
   name: 'Voice Agent',
   instructions: 'You are a voice assistant that can help users with their tasks.',
-  model: 'openai/gpt-5.4',
+  model: 'openai/gpt-5.5',
   voice: new OpenAIVoice(),
 })
 
@@ -375,7 +375,7 @@ const voiceAgent = new Agent({
   id: 'voice-agent',
   name: 'Voice Agent',
   instructions: 'You are a voice assistant that can help users with their tasks.',
-  model: 'openai/gpt-5.4',
+  model: 'openai/gpt-5.5',
   voice: new AzureVoice(),
 })
 
@@ -403,7 +403,7 @@ const voiceAgent = new Agent({
   id: 'voice-agent',
   name: 'Voice Agent',
   instructions: 'You are a voice assistant that can help users with their tasks.',
-  model: 'openai/gpt-5.4',
+  model: 'openai/gpt-5.5',
   voice: new ElevenLabsVoice(),
 })
 
@@ -431,7 +431,7 @@ const voiceAgent = new Agent({
   id: 'voice-agent',
   name: 'Voice Agent',
   instructions: 'You are a voice assistant that can help users with their tasks.',
-  model: 'openai/gpt-5.4',
+  model: 'openai/gpt-5.5',
   voice: new GoogleVoice(),
 })
 
@@ -459,7 +459,7 @@ const voiceAgent = new Agent({
   id: 'voice-agent',
   name: 'Voice Agent',
   instructions: 'You are a voice assistant that can help users with their tasks.',
-  model: 'openai/gpt-5.4',
+  model: 'openai/gpt-5.5',
   voice: new CloudflareVoice(),
 })
 
@@ -487,7 +487,7 @@ const voiceAgent = new Agent({
   id: 'voice-agent',
   name: 'Voice Agent',
   instructions: 'You are a voice assistant that can help users with their tasks.',
-  model: 'openai/gpt-5.4',
+  model: 'openai/gpt-5.5',
   voice: new DeepgramVoice(),
 })
 
@@ -515,7 +515,7 @@ const voiceAgent = new Agent({
   id: 'voice-agent',
   name: 'Voice Agent',
   instructions: 'You are a voice assistant that can help users with their tasks.',
-  model: 'openai/gpt-5.4',
+  model: 'openai/gpt-5.5',
   voice: new InworldVoice(),
 })
 
@@ -543,7 +543,7 @@ const voiceAgent = new Agent({
   id: 'voice-agent',
   name: 'Voice Agent',
   instructions: 'You are a voice assistant that can help users with their tasks.',
-  model: 'openai/gpt-5.4',
+  model: 'openai/gpt-5.5',
   voice: new SarvamVoice(),
 })
 
@@ -575,7 +575,7 @@ const voiceAgent = new Agent({
   id: 'voice-agent',
   name: 'Voice Agent',
   instructions: 'You are a voice assistant that can help users with their tasks.',
-  model: 'openai/gpt-5.4',
+  model: 'openai/gpt-5.5',
   voice: new OpenAIRealtimeVoice(),
 })
 
@@ -605,7 +605,7 @@ const voiceAgent = new Agent({
   id: 'voice-agent',
   name: 'Voice Agent',
   instructions: 'You are a voice assistant that can help users with their tasks.',
-  model: 'openai/gpt-5.4',
+  model: 'openai/gpt-5.5',
   voice: new GeminiLiveVoice({
     // Live API mode
     apiKey: process.env.GOOGLE_API_KEY,
@@ -654,7 +654,7 @@ const voiceAgent = new Agent({
   id: 'voice-agent',
   name: 'Voice Agent',
   instructions: 'You are a voice assistant that can help users with their tasks.',
-  model: 'openai/gpt-5.4',
+  model: 'openai/gpt-5.5',
   voice: new NovaSonicVoice({
     region: 'us-east-1',
     speaker: 'matthew',
@@ -697,7 +697,7 @@ const voiceAgent = new Agent({
   id: 'voice-agent',
   name: 'Voice Agent',
   instructions: 'You are a voice assistant that can help users with their tasks.',
-  model: 'openai/gpt-5.4',
+  model: 'openai/gpt-5.5',
   voice: new InworldRealtimeVoice({
     apiKey: process.env.INWORLD_API_KEY,
     model: 'inworld/models/gemma-4-26b-a4b-it',
@@ -1132,7 +1132,7 @@ const voiceAgent = new Agent({
   id: 'aisdk-voice-agent',
   name: 'AI SDK Voice Agent',
   instructions: 'You are a helpful assistant with voice capabilities.',
-  model: 'openai/gpt-5.4',
+  model: 'openai/gpt-5.5',
   voice,
 })
 ```

package/dist/docs/references/docs-voice-speech-to-speech.md

@@ -32,7 +32,7 @@ const agent = new Agent({
   id: 'agent',
   name: 'OpenAI Realtime Agent',
   instructions: `You are a helpful assistant with real-time voice capabilities.`,
-  model: 'openai/gpt-5.4',
+  model: 'openai/gpt-5.5',
   voice: new OpenAIRealtimeVoice(),
 })
 
@@ -66,7 +66,7 @@ const agent = new Agent({
   name: 'Gemini Live Agent',
   instructions: 'You are a helpful assistant with real-time voice capabilities.',
   // Model used for text generation; voice provider handles realtime audio
-  model: 'openai/gpt-5.4',
+  model: 'openai/gpt-5.5',
   voice: new GeminiLiveVoice({
     apiKey: process.env.GOOGLE_API_KEY,
     model: 'gemini-2.0-flash-exp',
@@ -113,7 +113,7 @@ const agent = new Agent({
   name: 'Nova Sonic Agent',
   instructions: 'You are a helpful assistant with real-time voice capabilities.',
   // Model used for text generation; voice provider handles realtime audio
-  model: 'openai/gpt-5.4',
+  model: 'openai/gpt-5.5',
   voice: new NovaSonicVoice({
     region: 'us-east-1',
     speaker: 'matthew',
@@ -157,7 +157,7 @@ const agent = new Agent({
   name: 'Inworld Realtime Agent',
   instructions: 'You are a helpful assistant with real-time voice capabilities.',
   // Model used for text generation; voice provider handles realtime audio
-  model: 'openai/gpt-5.4',
+  model: 'openai/gpt-5.5',
   voice: new InworldRealtimeVoice({
     apiKey: process.env.INWORLD_API_KEY,
     model: 'inworld/models/gemma-4-26b-a4b-it',

package/dist/docs/references/reference-voice-google-gemini-live.md

@@ -242,7 +242,9 @@ The GeminiLiveVoice class emits the following events:
 
 **speaking** (`event`): Emitted with audio metadata. Callback receives { audioData?: Int16Array, sampleRate?: number }.
 
-**writing** (`event`): Emitted when transcribed text is available. Callback receives { text: string, role: 'assistant' | 'user' }.
+**writing** (`event`): Emitted when transcribed text is available. Callback receives { text: string, role: 'assistant' | 'user' }. On native-audio models the assistant transcript is driven by the server's \`output\_audio\_transcription\` channel rather than \`modelTurn.parts.text\`.
+
+**thinking** (`event`): Emitted on native-audio models with the model's chain-of-thought / reasoning text from \`modelTurn.parts.text\`. Callback receives { text: string }. Does not fire on non-native-audio models, where \`modelTurn.parts.text\` is the spoken response and is emitted as \`writing\` instead.
 
 **session** (`event`): Emitted on session state changes. Callback receives { state: 'connecting' | 'connected' | 'disconnected' | 'disconnecting' | 'updated', config?: object }.
 
@@ -254,7 +256,18 @@ The GeminiLiveVoice class emits the following events:
 
 **error** (`event`): Emitted when an error occurs. Callback receives { message: string, code?: string, details?: unknown }.
 
-**interrupt** (`event`): Interrupt events. Callback receives { type: 'user' | 'model', timestamp: number }.
+**interrupt** (`event`): Emitted on barge-in when the user starts speaking over an in-flight model response. The server cancels any further audio for the current turn. Callback receives { type: 'user', timestamp: number }.
+
+## Native-audio behavior
+
+Native-audio Gemini Live models — any model whose ID contains `native-audio`, such as `gemini-2.5-flash-native-audio-preview-12-2025` — split text output across two channels:
+
+- The model's spoken reply is delivered as audio plus an `output_audio_transcription` transcript and surfaced as `writing` with `role: 'assistant'`.
+- The model's internal reasoning is delivered as `modelTurn.parts.text` and surfaced as `thinking`.
+
+On non-native-audio models there is no `output_audio_transcription` channel, so `modelTurn.parts.text` is the spoken response itself and is emitted as `writing`; the `thinking` event does not fire.
+
+Input transcription, output transcription, and barge-in detection (`realtime_input_config.activity_handling = 'START_OF_ACTIVITY_INTERRUPTS'`) are enabled automatically in the setup payload — no extra configuration is required.
 
 ## Available models
 

package/dist/index.cjs

@@ -1529,6 +1529,10 @@ var GeminiLiveVoice = class _GeminiLiveVoice extends MastraVoice {
   requestContext;
   // Store the configuration options
   options;
+  // Accumulates assistant text across `serverContent` frames for the current
+  // turn. Live API streams responses over many frames; we aggregate here and
+  // flush to context history once on `turnComplete`.
+  pendingAssistantResponse = "";
   /**
    * Normalize configuration to ensure proper VoiceConfig format
    * Handles backward compatibility with direct GeminiLiveVoiceConfig
@@ -2483,15 +2487,38 @@ var GeminiLiveVoice = class _GeminiLiveVoice extends MastraVoice {
     if (!data) {
       return;
     }
-    let assistantResponse = "";
+    if (data.interrupted) {
+      this.log("Model response interrupted by user activity");
+      this.audioStreamManager.cleanupSpeakerStreams();
+      this.pendingAssistantResponse = "";
+      this.emit("interrupt", { type: "user", timestamp: Date.now() });
+    }
+    if (data.inputTranscription?.text) {
+      this.emit("writing", {
+        text: data.inputTranscription.text,
+        role: "user"
+      });
+    }
+    if (data.outputTranscription?.text) {
+      this.pendingAssistantResponse += data.outputTranscription.text;
+      this.emit("writing", {
+        text: data.outputTranscription.text,
+        role: "assistant"
+      });
+    }
+    const nativeAudio = this.isNativeAudioModel();
     if (data.modelTurn?.parts) {
       for (const part of data.modelTurn.parts) {
         if (part.text) {
-          assistantResponse += part.text;
-          this.emit("writing", {
-            text: part.text,
-            role: "assistant"
-          });
+          if (nativeAudio) {
+            this.emit("thinking", { text: part.text });
+          } else {
+            this.pendingAssistantResponse += part.text;
+            this.emit("writing", {
+              text: part.text,
+              role: "assistant"
+            });
+          }
         }
         if (part.functionCall) {
           this.log("Found function call in serverContent.modelTurn.parts", part.functionCall);
@@ -2560,11 +2587,12 @@ var GeminiLiveVoice = class _GeminiLiveVoice extends MastraVoice {
         }
       }
     }
-    if (assistantResponse.trim()) {
-      this.addToContext("assistant", assistantResponse);
-    }
     if (data.turnComplete) {
       this.log("Turn completed");
+      if (this.pendingAssistantResponse.trim()) {
+        this.addToContext("assistant", this.pendingAssistantResponse);
+      }
+      this.pendingAssistantResponse = "";
       this.audioStreamManager.cleanupSpeakerStreams();
       this.emit("turnComplete", {
         timestamp: Date.now()
@@ -2717,6 +2745,27 @@ var GeminiLiveVoice = class _GeminiLiveVoice extends MastraVoice {
   getVertexLocation() {
     return this.options.location?.trim() || "us-central1";
   }
+  /**
+   * Whether the active model is a Gemini Live "native-audio" variant.
+   *
+   * Native-audio models emit a different `serverContent.modelTurn.parts.text`
+   * stream than their half-cascade siblings: on native-audio, that text is the
+   * model's internal reasoning (chain-of-thought), and the *spoken* response
+   * arrives separately via `serverContent.outputTranscription.text`. On
+   * non-native-audio models there is no `outputTranscription` channel, and
+   * `modelTurn.parts.text` is the spoken response.
+   *
+   * Used to decide whether `modelTurn.parts.text` should be emitted as
+   * `thinking` (native-audio) or `writing` (non-native-audio). All native-audio
+   * model IDs in `GeminiVoiceModel` contain the literal substring
+   * `native-audio`, so a substring check is sufficient and forward-compatible
+   * with new variants that follow the same naming convention.
+   * @private
+   */
+  isNativeAudioModel() {
+    const model = this.options.model ?? DEFAULT_MODEL;
+    return model.includes("native-audio");
+  }
   /**
    * Resolve the correct model identifier for Gemini API or Vertex AI
    * @private
@@ -2758,7 +2807,20 @@ var GeminiLiveVoice = class _GeminiLiveVoice extends MastraVoice {
     const setupMessage = {
       setup: {
         model: this.resolveModelIdentifier(),
-        generation_config: generationConfig
+        generation_config: generationConfig,
+        // Transcription is on by default — matches the pattern in @mastra/voice-openai-realtime,
+        // @mastra/voice-xai-realtime, and @mastra/voice-inworld where realtime sessions
+        // unconditionally enable STT in `connect()`. On native-audio models this is the ONLY way
+        // to receive the spoken response as text (`modelTurn.parts.text` carries reasoning, not
+        // speech), so without these flags the assistant's words are silently dropped client-side.
+        input_audio_transcription: {},
+        output_audio_transcription: {},
+        // Activity-based interrupts surface barge-in as `serverContent.interrupted = true` and
+        // cancel the in-flight model response. This is the only way to wire up the `interrupt`
+        // event declared in `GeminiLiveEventMap`.
+        realtime_input_config: {
+          activity_handling: "START_OF_ACTIVITY_INTERRUPTS"
+        }
       }
     };
     if (this.options.instructions) {