Claude/improve app quality oq agb

README.md

@@ -16,6 +16,41 @@
 </table>
 
 
+## 🤖 Mimo Bot — voice companion (work in progress)
+
+A small ESP32-S3 device acts as a mic + speaker for the phone; the phone runs
+the full Whisper → LLM → TTS pipeline using LLM Hub's existing inference core.
+Nothing leaves the pair of devices.
+
+```
+┌───────── ESP32-S3 ─────────┐      ┌───── Phone (LLM Hub) ─────┐
+│ mic → Opus → BLE   ────────┼──────┼► Whisper → LLM → TTS      │
+│ spk ← Opus ← BLE   ────────┼──────┼◄ Opus ← TTS (Kokoro/Sys)  │
+└────────────────────────────┘      └───────────────────────────┘
+```
+
+Status: **scaffolding only.** The transport, audio codec, Whisper, TTS, and
+pipeline modules are stubs with `TODO(...)` markers and a locked wire protocol
+so firmware and apps can be built in parallel without re-negotiating.
+
+- `docs/mimobot-protocol.md` — BLE service + control JSON schema
+- `android/app/src/main/java/com/llmhub/llmhub/mimobot/` — Android pipeline
+  (`transport/`, `audio/`, `speech/`, `pipeline/`)
+- `ios/LLMHub/Sources/LLMHub/MimoBot/` — iOS pipeline (same shape)
+- `firmware/` — ESP-IDF v5.2+ skeleton for the ESP32-S3
+
+The pipeline reuses the existing `InferenceService` / `LLMBackend`,
+`RagService`, and web-search services unchanged — voice is purely additive.
+
+TTS is pluggable via a `Tts` / `TTS` interface; two implementations are
+stubbed out:
+- **SystemTts / SystemTTS** — free, ships with the OS, decent quality
+- **KokoroTts / KokoroTTS** — Kokoro-82M via ONNX Runtime (already a
+  dependency), higher quality, one ~160 MB download
+
+See `firmware/README.md` for hardware and toolchain.
+
+
 ## 📸 Screenshots
 
 <div style="display:flex;gap:12px;flex-wrap:wrap;align-items:flex-start;">

android/app/build.gradle.kts

@@ -153,6 +153,20 @@ android {
     }
 
     // Removed externalNativeBuild - now using MediaPipe instead of native llama.cpp
+
+    // Mimo Bot espeak-ng JNI bridge. Only enabled when the user has run
+    // scripts/build_espeak_android.sh and the headers landed under
+    // src/main/cpp/espeak-ng/include/. Skipping the block on fresh checkouts
+    // means AGP doesn't require an NDK install for users who only want the
+    // bundled DictionaryG2P fallback.
+    if (file("src/main/cpp/espeak-ng/include/speak_lib.h").exists()) {
+        externalNativeBuild {
+            cmake {
+                path = file("src/main/cpp/CMakeLists.txt")
+                version = "3.22.1"
+            }
+        }
+    }
 }
 
 // The Image Generator requires full protobuf-java, not the lite version

android/app/src/main/AndroidManifest.xml

@@ -15,6 +15,10 @@
         <intent>
             <action android:name="android.intent.action.TTS_SERVICE" />
         </intent>
+        <!-- Required on Android 11+ so SpeechRecognizer can bind to the on-device recognizer. -->
+        <intent>
+            <action android:name="android.speech.RecognitionService" />
+        </intent>
     </queries>
     <!-- end -->
 

android/app/src/main/cpp/CMakeLists.txt

@@ -0,0 +1,39 @@
+# JNI bridge that exposes espeak-ng's text→IPA function to Kotlin.
+#
+# Layout assumed (populated by scripts/build_espeak_android.sh):
+#   src/main/cpp/
+#     espeak_jni.cpp          — this directory
+#     espeak-ng/include/      — espeak-ng public headers (speak_lib.h)
+#   src/main/jniLibs/<abi>/
+#     libespeak-ng.so         — cross-compiled espeak-ng runtime per ABI
+#   src/main/assets/espeak-ng-data/
+#     ...                     — espeak-ng's data files (voices, phoneme tables)
+#
+# If espeak-ng/include/ doesn't exist (e.g. fresh checkout), this file does
+# nothing — the Kotlin layer's `EspeakG2P.tryLoad()` will return null and the
+# pipeline falls back to the bundled DictionaryG2P.
+
+cmake_minimum_required(VERSION 3.18.1)
+project(mimo_espeak_jni LANGUAGES CXX)
+
+set(CMAKE_CXX_STANDARD 17)
+
+set(ESPEAK_INCLUDE_DIR "${CMAKE_CURRENT_SOURCE_DIR}/espeak-ng/include")
+
+if(NOT EXISTS "${ESPEAK_INCLUDE_DIR}/speak_lib.h")
+    message(WARNING "espeak-ng headers not found at ${ESPEAK_INCLUDE_DIR}. "
+                    "Run scripts/build_espeak_android.sh to set them up. "
+                    "Skipping espeak JNI build.")
+    return()
+endif()
+
+# Locate the per-ABI prebuilt libespeak-ng.so that the build script dropped
+# under src/main/jniLibs/<abi>/. Gradle automatically packages everything in
+# jniLibs into the APK/AAB, so we just declare the import target here.
+add_library(espeak-ng SHARED IMPORTED)
+set_target_properties(espeak-ng PROPERTIES
+    IMPORTED_LOCATION "${CMAKE_SOURCE_DIR}/../jniLibs/${ANDROID_ABI}/libespeak-ng.so")
+
+add_library(espeak-jni SHARED espeak_jni.cpp)
+target_include_directories(espeak-jni PRIVATE "${ESPEAK_INCLUDE_DIR}")
+target_link_libraries(espeak-jni PRIVATE espeak-ng log)

android/app/src/main/cpp/espeak_jni.cpp

@@ -0,0 +1,95 @@
+// Mimo bot — JNI bridge from Kotlin to espeak-ng for text→IPA phonemization.
+//
+// Symbols exported here are loaded by EspeakG2P.kt via System.loadLibrary.
+// All entry points are namespace `Java_com_llmhub_llmhub_mimobot_speech_kokoro_EspeakG2P_*`.
+//
+// espeak-ng API reference: speak_lib.h
+//   espeak_Initialize(AUDIO_OUTPUT_SYNCHRONOUS, 0, dataPath, 0)
+//   espeak_SetVoiceByName("en-us")
+//   espeak_TextToPhonemes(&inputBuf, espeakCHARS_AUTO,
+//                         espeakPHONEMES_IPA | espeakPHONEMES_SHOW)
+//
+// We deliberately keep this thin: tokenisation / vocabulary mapping happens in
+// Kotlin (KokoroVocab) so we can swap espeak for another phonemizer (e.g.
+// misaki, custom CMU dict) without touching native code.
+
+#include <jni.h>
+#include <cstring>
+#include <string>
+#include <android/log.h>
+
+extern "C" {
+#include "speak_lib.h"
+}
+
+#define LOG_TAG "MimoEspeakJni"
+#define LOGI(...) __android_log_print(ANDROID_LOG_INFO, LOG_TAG, __VA_ARGS__)
+#define LOGW(...) __android_log_print(ANDROID_LOG_WARN, LOG_TAG, __VA_ARGS__)
+
+static bool g_initialised = false;
+
+extern "C" JNIEXPORT jint JNICALL
+Java_com_llmhub_llmhub_mimobot_speech_kokoro_EspeakG2P_nativeInit(
+    JNIEnv* env, jclass /*clazz*/, jstring jDataPath) {
+    if (g_initialised) return 0;
+    const char* dataPath = env->GetStringUTFChars(jDataPath, nullptr);
+    // AUDIO_OUTPUT_SYNCHRONOUS with no buffer — we never play audio, only ask
+    // for phonemes. Returning sample rate or -1 on failure.
+    int rate = espeak_Initialize(AUDIO_OUTPUT_SYNCHRONOUS, 0, dataPath, 0);
+    env->ReleaseStringUTFChars(jDataPath, dataPath);
+    if (rate < 0) {
+        LOGW("espeak_Initialize failed (rate=%d)", rate);
+        return rate;
+    }
+    g_initialised = true;
+    LOGI("espeak-ng initialised (rate=%d)", rate);
+    return 0;
+}
+
+extern "C" JNIEXPORT jint JNICALL
+Java_com_llmhub_llmhub_mimobot_speech_kokoro_EspeakG2P_nativeSetVoice(
+    JNIEnv* env, jclass /*clazz*/, jstring jVoiceName) {
+    const char* voice = env->GetStringUTFChars(jVoiceName, nullptr);
+    espeak_ERROR rc = espeak_SetVoiceByName(voice);
+    env->ReleaseStringUTFChars(jVoiceName, voice);
+    if (rc != EE_OK) {
+        LOGW("espeak_SetVoiceByName(%s) failed rc=%d", voice, (int)rc);
+        return -1;
+    }
+    return 0;
+}
+
+extern "C" JNIEXPORT jstring JNICALL
+Java_com_llmhub_llmhub_mimobot_speech_kokoro_EspeakG2P_nativePhonemize(
+    JNIEnv* env, jclass /*clazz*/, jstring jText) {
+    if (!g_initialised) return env->NewStringUTF("");
+    const char* utf = env->GetStringUTFChars(jText, nullptr);
+
+    // espeak_TextToPhonemes:
+    //   - mode bits: 0x02 = IPA, 0x01 = print stress markers (we want both).
+    //     espeakPHONEMES_IPA  = 0x02
+    //     espeakPHONEMES_SHOW = 0x01 (returned, not printed when no FILE*)
+    //   - input is consumed; we pass &input which the function advances.
+    const void* input = utf;
+    std::string out;
+    while (input != nullptr) {
+        const char* phonemes = espeak_TextToPhonemes(
+            &input,
+            /*textmode*/ espeakCHARS_AUTO,
+            /*phonememode*/ 0x02);  // IPA, no stress separators
+        if (phonemes == nullptr) break;
+        if (!out.empty() && out.back() != ' ') out.push_back(' ');
+        out.append(phonemes);
+    }
+
+    env->ReleaseStringUTFChars(jText, utf);
+    return env->NewStringUTF(out.c_str());
+}
+
+extern "C" JNIEXPORT void JNICALL
+Java_com_llmhub_llmhub_mimobot_speech_kokoro_EspeakG2P_nativeShutdown(
+    JNIEnv* /*env*/, jclass /*clazz*/) {
+    if (!g_initialised) return;
+    espeak_Terminate();
+    g_initialised = false;
+}

android/app/src/main/java/com/llmhub/llmhub/data/ModelData.kt

@@ -57,6 +57,76 @@ object DeviceInfo {
 
 object ModelData {
     val models = listOf(
+        // Gemma-3 270M — smallest Gemma 3, great for fast voice replies on the
+        // Mimo Bot test screen. Multiple quantization × KV-cache variants
+        // following the same litert-community filename convention as Gemma-3 1B.
+        // If any URL 404s, the rest of the entry stays correct — just edit the URL.
+        LLMModel(
+            name = "Gemma-3 270M (INT4, 1k)",
+            description = "Google Gemma-3 270M, INT4 quantization, 1k context. Smallest + fastest. Ideal for low-latency voice replies. (~150 MB)",
+            url = "https://huggingface.co/litert-community/Gemma3-270M-IT/resolve/main/Gemma3-270M-IT_multi-prefill-seq_q4_ekv1024.task?download=true",
+            category = "text",
+            sizeBytes = 157286400L, // ~150 MB
+            source = "Google via LiteRT Community",
+            supportsVision = false,
+            supportsGpu = true,
+            requirements = ModelRequirements(minRamGB = 1, recommendedRamGB = 2),
+            contextWindowSize = 1024,
+            modelFormat = "task"
+        ),
+        LLMModel(
+            name = "Gemma-3 270M (INT4, 2k)",
+            description = "Google Gemma-3 270M, INT4, 2k context. Slightly more memory for longer turns. (~160 MB)",
+            url = "https://huggingface.co/litert-community/Gemma3-270M-IT/resolve/main/Gemma3-270M-IT_multi-prefill-seq_q4_ekv2048.task?download=true",
+            category = "text",
+            sizeBytes = 167772160L, // ~160 MB
+            source = "Google via LiteRT Community",
+            supportsVision = false,
+            supportsGpu = true,
+            requirements = ModelRequirements(minRamGB = 1, recommendedRamGB = 2),
+            contextWindowSize = 2048,
+            modelFormat = "task"
+        ),
+        LLMModel(
+            name = "Gemma-3 270M (INT8, 1k)",
+            description = "Google Gemma-3 270M, INT8 quantization, 1k context. Higher quality, still fast. (~270 MB)",
+            url = "https://huggingface.co/litert-community/Gemma3-270M-IT/resolve/main/Gemma3-270M-IT_multi-prefill-seq_q8_ekv1024.task?download=true",
+            category = "text",
+            sizeBytes = 283115520L, // ~270 MB
+            source = "Google via LiteRT Community",
+            supportsVision = false,
+            supportsGpu = true,
+            requirements = ModelRequirements(minRamGB = 1, recommendedRamGB = 2),
+            contextWindowSize = 1024,
+            modelFormat = "task"
+        ),
+        LLMModel(
+            name = "Gemma-3 270M (INT8, 2k)",
+            description = "Google Gemma-3 270M, INT8, 2k context. Best balance of quality + context for the Mimo Bot loop. (~280 MB)",
+            url = "https://huggingface.co/litert-community/Gemma3-270M-IT/resolve/main/Gemma3-270M-IT_multi-prefill-seq_q8_ekv2048.task?download=true",
+            category = "text",
+            sizeBytes = 293601280L, // ~280 MB
+            source = "Google via LiteRT Community",
+            supportsVision = false,
+            supportsGpu = true,
+            requirements = ModelRequirements(minRamGB = 1, recommendedRamGB = 2),
+            contextWindowSize = 2048,
+            modelFormat = "task"
+        ),
+        LLMModel(
+            name = "Gemma-3 270M (INT8, 4k)",
+            description = "Google Gemma-3 270M, INT8, 4k context. For longer multi-turn voice sessions. (~290 MB)",
+            url = "https://huggingface.co/litert-community/Gemma3-270M-IT/resolve/main/Gemma3-270M-IT_multi-prefill-seq_q8_ekv4096.task?download=true",
+            category = "text",
+            sizeBytes = 304087040L, // ~290 MB
+            source = "Google via LiteRT Community",
+            supportsVision = false,
+            supportsGpu = true,
+            requirements = ModelRequirements(minRamGB = 2, recommendedRamGB = 3),
+            contextWindowSize = 4096,
+            modelFormat = "task"
+        ),
+
         // Gemma-3 1B Models
         LLMModel(
             name = "Gemma-3 1B (INT4, 2k)",

android/app/src/main/java/com/llmhub/llmhub/mimobot/MimoBotIds.kt

@@ -0,0 +1,23 @@
+package com.llmhub.llmhub.mimobot
+
+import java.util.UUID
+
+/**
+ * Canonical identifiers shared across the Mimo bot stack.
+ * Must stay in sync with docs/mimobot-protocol.md and firmware/main/ble_transport.c.
+ */
+object MimoBotIds {
+    val SERVICE: UUID    = UUID.fromString("6d696d6f-b07d-4e13-9e88-000000000001")
+    val AUDIO_UP: UUID   = UUID.fromString("6d696d6f-b07d-4e13-9e88-000000000010")
+    val AUDIO_DOWN: UUID = UUID.fromString("6d696d6f-b07d-4e13-9e88-000000000011")
+    val CONTROL: UUID    = UUID.fromString("6d696d6f-b07d-4e13-9e88-000000000012")
+
+    const val ADV_NAME_PREFIX = "mimobot-"
+
+    // Audio format constants (see docs/mimobot-protocol.md)
+    const val SAMPLE_RATE_HZ = 16_000
+    const val FRAME_MS = 20
+    const val FRAME_SAMPLES = SAMPLE_RATE_HZ * FRAME_MS / 1000  // 320
+    const val CHANNELS = 1
+    const val OPUS_BITRATE = 24_000
+}

android/app/src/main/java/com/llmhub/llmhub/mimobot/audio/AudioIO.kt

@@ -0,0 +1,23 @@
+package com.llmhub.llmhub.mimobot.audio
+
+import kotlinx.coroutines.flow.Flow
+
+/**
+ * Abstractions that let [com.llmhub.llmhub.mimobot.pipeline.VoicePipeline] run
+ * against either the phone's own mic+speaker (local dev mode) or a BLE
+ * transport with Opus codec (real device mode).
+ *
+ * Contract: all frames are 16 kHz mono Int16 PCM, 320 samples (20 ms) each.
+ */
+interface AudioSource {
+    /** Start capturing; flow completes when [stop] is called. */
+    fun frames(): Flow<ShortArray>
+    fun stop()
+}
+
+interface AudioSink {
+    /** Must be called before [write]. Idempotent. */
+    fun start()
+    suspend fun write(frame: ShortArray)
+    fun stop()
+}