# Phase 7 - Implémentation STT Modulaire

**Date de création** : 2025-11-29
**Objectif** : Architecture STT complète avec support multi-engines (Vosk, PocketSphinx, Whisper)
**Nom de l'assistant** : Celuna (anciennement AISSIA)

---

## Vue d'Ensemble

### Objectifs

1. **Architecture modulaire** : Interface `ISTTEngine` avec 4 implémentations
2. **Service STT** : Layer `ISTTService` pour abstraction business logic
3. **Dual Mode** : Passive (keyword spotting) + Active (transcription complète)
4. **Coût optimisé** : Local par défaut, Whisper API en fallback optionnel

---

## Architecture Cible

```
┌─────────────────────────────────────────────────────────┐
│                    VoiceService                          │
│  - Gère TTS (EspeakTTSEngine)                           │
│  - Gère STT via ISTTService                             │
│  - Pub/sub IIO (voice:speak, voice:listen, etc.)       │
└─────────────────────────────────────────────────────────┘
                           │
                           ▼
┌─────────────────────────────────────────────────────────┐
│                    ISTTService                           │
│  - Interface service STT                                │
│  - Gère mode passive/active                             │
│  - Switch engines selon config                          │
│  - Fallback automatique                                 │
└─────────────────────────────────────────────────────────┘
                           │
                           ▼
┌─────────────────────────────────────────────────────────┐
│                  STTEngineFactory                        │
│  - create(type, config) → unique_ptr<ISTTEngine>        │
└─────────────────────────────────────────────────────────┘
                           │
          ┌────────────────┼────────────────┬──────────────┐
          ▼                ▼                ▼              ▼
   ┌──────────┐   ┌──────────────┐  ┌─────────────┐  ┌──────────────┐
   │  Vosk    │   │ PocketSphinx │  │ WhisperCpp  │  │ WhisperAPI   │
   │  Engine  │   │   Engine     │  │   Engine    │  │   Engine     │
   └──────────┘   └──────────────┘  └─────────────┘  └──────────────┘
   Local          Local (keywords)   Local (précis)   Remote (payant)
   50MB model     Léger ~10MB        75-142MB          API OpenAI
```

---

## Phase 7.1 - Service Layer (ISTTService)

### Objectif

Créer une couche service qui abstrait la complexité des engines STT et gère :
- Mode passive/active
- Switching d'engines
- Fallback automatique
- Gestion erreurs

### Fichiers à créer

#### 1. `src/services/ISTTService.hpp`

**Interface du service STT**

```cpp
#pragma once

#include <string>
#include <vector>
#include <functional>
#include <memory>

namespace aissia {

enum class STTMode {
    PASSIVE,  // Keyword spotting (économe)
    ACTIVE    // Full transcription
};

enum class STTEngineType {
    VOSK,
    POCKETSPHINX,
    WHISPER_CPP,
    WHISPER_API,
    AUTO  // Factory choisit
};

/**
 * @brief Callback pour résultats transcription
 */
using TranscriptionCallback = std::function<void(const std::string& text, STTMode mode)>;

/**
 * @brief Callback pour détection keyword
 */
using KeywordCallback = std::function<void(const std::string& keyword)>;

/**
 * @brief Interface service STT
 */
class ISTTService {
public:
    virtual ~ISTTService() = default;

    /**
     * @brief Démarre le service STT
     */
    virtual bool start() = 0;

    /**
     * @brief Arrête le service STT
     */
    virtual void stop() = 0;

    /**
     * @brief Change le mode STT
     */
    virtual void setMode(STTMode mode) = 0;

    /**
     * @brief Obtient le mode actuel
     */
    virtual STTMode getMode() const = 0;

    /**
     * @brief Transcrit un fichier audio
     */
    virtual std::string transcribeFile(const std::string& filePath) = 0;

    /**
     * @brief Transcrit des données audio PCM
     */
    virtual std::string transcribe(const std::vector<float>& audioData) = 0;

    /**
     * @brief Active l'écoute en streaming (temps réel)
     */
    virtual void startListening(TranscriptionCallback onTranscription,
                                KeywordCallback onKeyword) = 0;

    /**
     * @brief Arrête l'écoute streaming
     */
    virtual void stopListening() = 0;

    /**
     * @brief Configure la langue
     */
    virtual void setLanguage(const std::string& language) = 0;

    /**
     * @brief Vérifie si le service est disponible
     */
    virtual bool isAvailable() const = 0;

    /**
     * @brief Obtient le nom de l'engine actuel
     */
    virtual std::string getCurrentEngine() const = 0;
};

} // namespace aissia
```

**Estimation** : 50 lignes

---

#### 2. `src/services/STTService.hpp` + `.cpp`

**Implémentation du service STT**

**Features** :
- Gère 2 engines : 1 pour passive (PocketSphinx), 1 pour active (Vosk/Whisper)
- Switch automatique passive → active sur keyword
- Timeout active → passive (30s sans parole)
- Fallback vers Whisper API si engine local fail
- Thread d'écoute microphone (via PortAudio ou ALSA)

**Pseudo-code** :

```cpp
class STTService : public ISTTService {
private:
    std::unique_ptr<ISTTEngine> m_passiveEngine;  // PocketSphinx
    std::unique_ptr<ISTTEngine> m_activeEngine;   // Vosk/Whisper
    std::unique_ptr<ISTTEngine> m_fallbackEngine; // WhisperAPI

    STTMode m_currentMode = STTMode::PASSIVE;

    std::thread m_listenThread;
    std::atomic<bool> m_listening{false};

    TranscriptionCallback m_onTranscription;
    KeywordCallback m_onKeyword;

    std::chrono::steady_clock::time_point m_lastActivity;

public:
    bool start() override {
        // Load engines from config
        m_passiveEngine = STTEngineFactory::create("pocketsphinx", config);
        m_activeEngine = STTEngineFactory::create("vosk", config);
        m_fallbackEngine = STTEngineFactory::create("whisper-api", config);

        return m_passiveEngine && m_activeEngine;
    }

    void startListening(TranscriptionCallback onTranscription,
                       KeywordCallback onKeyword) override {
        m_onTranscription = onTranscription;
        m_onKeyword = onKeyword;

        m_listening = true;
        m_listenThread = std::thread([this]() {
            listenLoop();
        });
    }

private:
    void listenLoop() {
        // Ouvrir microphone (PortAudio)
        // Boucle infinie :
        //   - Si PASSIVE : use m_passiveEngine (keywords only)
        //     - Si keyword détecté → setMode(ACTIVE) + callback
        //   - Si ACTIVE : use m_activeEngine (full transcription)
        //     - Transcrit en temps réel
        //     - Si timeout 30s → setMode(PASSIVE)
    }
};
```

**Estimation** : 300 lignes (service + thread microphone)

---

## Phase 7.2 - Engines STT

### Fichiers à modifier/créer

#### 1. `src/shared/audio/ISTTEngine.hpp` ✅ Existe

**Modifications** : Aucune (interface déjà bonne)

---

#### 2. `src/shared/audio/WhisperAPIEngine.hpp` ✅ Existe

**Modifications** : Aucune (déjà implémenté, sera utilisé comme fallback)

---

#### 3. `src/shared/audio/VoskSTTEngine.hpp` 🆕 À créer

**Vosk Speech Recognition**

**Dépendances** :
- `vosk` library (C++ bindings)
- Modèle français : `vosk-model-small-fr-0.22` (~50MB)

**Installation** :
```bash
# Linux
sudo apt install libvosk-dev

# Télécharger modèle FR
wget https://alphacephei.com/vosk/models/vosk-model-small-fr-0.22.zip
unzip vosk-model-small-fr-0.22.zip -d models/
```

**Implémentation** :

```cpp
#pragma once

#include "ISTTEngine.hpp"
#include <vosk_api.h>
#include <spdlog/spdlog.h>

namespace aissia {

class VoskSTTEngine : public ISTTEngine {
public:
    explicit VoskSTTEngine(const std::string& modelPath) {
        m_logger = spdlog::get("VoskSTT");
        if (!m_logger) {
            m_logger = spdlog::stdout_color_mt("VoskSTT");
        }

        // Load Vosk model
        m_model = vosk_model_new(modelPath.c_str());
        if (!m_model) {
            m_logger->error("Failed to load Vosk model: {}", modelPath);
            m_available = false;
            return;
        }

        // Create recognizer (16kHz, mono)
        m_recognizer = vosk_recognizer_new(m_model, 16000.0);
        m_available = true;

        m_logger->info("Vosk STT initialized: {}", modelPath);
    }

    ~VoskSTTEngine() override {
        if (m_recognizer) vosk_recognizer_free(m_recognizer);
        if (m_model) vosk_model_free(m_model);
    }

    std::string transcribe(const std::vector<float>& audioData) override {
        if (!m_available || audioData.empty()) return "";

        // Convert float to int16
        std::vector<int16_t> samples(audioData.size());
        for (size_t i = 0; i < audioData.size(); ++i) {
            samples[i] = static_cast<int16_t>(audioData[i] * 32767.0f);
        }

        // Feed audio to recognizer
        vosk_recognizer_accept_waveform(m_recognizer,
            reinterpret_cast<const char*>(samples.data()),
            samples.size() * sizeof(int16_t));

        // Get final result
        const char* result = vosk_recognizer_final_result(m_recognizer);

        // Parse JSON result: {"text": "transcription"}
        std::string text = parseVoskResult(result);

        m_logger->debug("Transcribed: {}", text);
        return text;
    }

    std::string transcribeFile(const std::string& filePath) override {
        // Load WAV file, convert to PCM, call transcribe()
        // (Implementation omitted for brevity)
    }

    void setLanguage(const std::string& language) override {
        // Vosk model is language-specific, can't change at runtime
    }

    bool isAvailable() const override { return m_available; }
    std::string getEngineName() const override { return "vosk"; }

private:
    VoskModel* m_model = nullptr;
    VoskRecognizer* m_recognizer = nullptr;
    bool m_available = false;
    std::shared_ptr<spdlog::logger> m_logger;

    std::string parseVoskResult(const char* json) {
        // Parse JSON: {"text": "bonjour"} → "bonjour"
        // Use nlohmann::json
    }
};

} // namespace aissia
```

**Estimation** : 200 lignes

---

#### 4. `src/shared/audio/PocketSphinxEngine.hpp` 🆕 À créer

**PocketSphinx Keyword Spotting**

**Dépendances** :
- `pocketsphinx` library
- Acoustic model (phonétique)

**Installation** :
```bash
sudo apt install pocketsphinx pocketsphinx-en-us
```

**Configuration Keywords** :
```
# keywords.txt
celuna /1e-40/
hey celuna /1e-50/
```

**Implémentation** :

```cpp
#pragma once

#include "ISTTEngine.hpp"
#include <pocketsphinx.h>
#include <spdlog/spdlog.h>

namespace aissia {

class PocketSphinxEngine : public ISTTEngine {
public:
    explicit PocketSphinxEngine(const std::vector<std::string>& keywords,
                                 const std::string& modelPath) {
        m_logger = spdlog::get("PocketSphinx");
        if (!m_logger) {
            m_logger = spdlog::stdout_color_mt("PocketSphinx");
        }

        // Create keyword file
        createKeywordFile(keywords);

        // Initialize PocketSphinx
        ps_config_t* config = ps_config_init(NULL);
        ps_config_set_str(config, "hmm", modelPath.c_str());
        ps_config_set_str(config, "kws", "/tmp/celuna_keywords.txt");
        ps_config_set_float(config, "kws_threshold", 1e-40);

        m_decoder = ps_init(config);
        m_available = (m_decoder != nullptr);

        if (m_available) {
            m_logger->info("PocketSphinx initialized for keyword spotting");
        }
    }

    ~PocketSphinxEngine() override {
        if (m_decoder) ps_free(m_decoder);
    }

    std::string transcribe(const std::vector<float>& audioData) override {
        if (!m_available || audioData.empty()) return "";

        // Convert to int16
        std::vector<int16_t> samples(audioData.size());
        for (size_t i = 0; i < audioData.size(); ++i) {
            samples[i] = static_cast<int16_t>(audioData[i] * 32767.0f);
        }

        // Process audio
        ps_start_utt(m_decoder);
        ps_process_raw(m_decoder, samples.data(), samples.size(), FALSE, FALSE);
        ps_end_utt(m_decoder);

        // Get keyword (if detected)
        const char* hyp = ps_get_hyp(m_decoder, nullptr);
        std::string keyword = (hyp ? hyp : "");

        if (!keyword.empty()) {
            m_logger->info("Keyword detected: {}", keyword);
        }

        return keyword;
    }

    std::string transcribeFile(const std::string& filePath) override {
        // Not used for keyword spotting (streaming only)
        return "";
    }

    void setLanguage(const std::string& language) override {}
    bool isAvailable() const override { return m_available; }
    std::string getEngineName() const override { return "pocketsphinx"; }

private:
    ps_decoder_t* m_decoder = nullptr;
    bool m_available = false;
    std::shared_ptr<spdlog::logger> m_logger;

    void createKeywordFile(const std::vector<std::string>& keywords) {
        std::ofstream file("/tmp/celuna_keywords.txt");
        for (const auto& kw : keywords) {
            file << kw << " /1e-40/\n";
        }
    }
};

} // namespace aissia
```

**Estimation** : 180 lignes

---

#### 5. `src/shared/audio/WhisperCppEngine.hpp` 🆕 À créer (OPTIONNEL)

**whisper.cpp - Local Whisper**

**Dépendances** :
- `whisper.cpp` (ggerganov)
- Modèle : `ggml-tiny.bin` (75MB) ou `ggml-base.bin` (142MB)

**Installation** :
```bash
git clone https://github.com/ggerganov/whisper.cpp external/whisper.cpp
cd external/whisper.cpp
make
./models/download-ggml-model.sh tiny
```

**Implémentation** : Similar à Vosk mais avec API whisper.cpp

**Estimation** : 250 lignes

**⚠️ Note** : Optionnel, à implémenter seulement si besoin haute précision locale

---

#### 6. `src/shared/audio/STTEngineFactory.cpp` 📝 Modifier

**Factory pattern pour créer engines**

```cpp
#include "STTEngineFactory.hpp"
#include "VoskSTTEngine.hpp"
#include "PocketSphinxEngine.hpp"
#include "WhisperCppEngine.hpp"
#include "WhisperAPIEngine.hpp"

namespace aissia {

std::unique_ptr<ISTTEngine> STTEngineFactory::create(
    const std::string& type,
    const nlohmann::json& config) {

    if (type == "vosk" || type == "auto") {
        std::string modelPath = config.value("model_path", "./models/vosk-model-small-fr-0.22");
        auto engine = std::make_unique<VoskSTTEngine>(modelPath);
        if (engine->isAvailable()) return engine;
    }

    if (type == "pocketsphinx") {
        std::vector<std::string> keywords = config.value("keywords", std::vector<std::string>{"celuna"});
        std::string modelPath = config.value("model_path", "/usr/share/pocketsphinx/model/en-us");
        auto engine = std::make_unique<PocketSphinxEngine>(keywords, modelPath);
        if (engine->isAvailable()) return engine;
    }

    if (type == "whisper-cpp") {
        std::string modelPath = config.value("model_path", "./models/ggml-tiny.bin");
        auto engine = std::make_unique<WhisperCppEngine>(modelPath);
        if (engine->isAvailable()) return engine;
    }

    if (type == "whisper-api") {
        std::string apiKey = std::getenv(config.value("api_key_env", "OPENAI_API_KEY").c_str());
        if (!apiKey.empty()) {
            return std::make_unique<WhisperAPIEngine>(apiKey);
        }
    }

    // Fallback: stub engine (no-op)
    return std::make_unique<StubSTTEngine>();
}

} // namespace aissia
```

**Estimation** : 80 lignes

---

## Phase 7.3 - Intégration VoiceService

### Fichier à modifier

#### `src/services/VoiceService.cpp`

**Modifications** :

1. **Remplacer implémentation directe par ISTTService**

**Avant** :
```cpp
// VoiceService gère directement WhisperAPIEngine
std::unique_ptr<WhisperAPIEngine> m_sttEngine;
```

**Après** :
```cpp
// VoiceService délègue à ISTTService
std::unique_ptr<ISTTService> m_sttService;
```

2. **Initialisation** :

```cpp
void VoiceService::initialize(const nlohmann::json& config) {
    // TTS (unchanged)
    m_ttsEngine = TTSEngineFactory::create();

    // STT (new)
    m_sttService = std::make_unique<STTService>(config["stt"]);
    m_sttService->start();

    // Setup callbacks
    m_sttService->startListening(
        [this](const std::string& text, STTMode mode) {
            handleTranscription(text, mode);
        },
        [this](const std::string& keyword) {
            handleKeyword(keyword);
        }
    );
}
```

3. **Handlers** :

```cpp
void VoiceService::handleKeyword(const std::string& keyword) {
    m_logger->info("Keyword detected: {}", keyword);

    // Publish keyword detection
    nlohmann::json event = {
        {"type", "keyword_detected"},
        {"keyword", keyword},
        {"timestamp", std::time(nullptr)}
    };
    m_io->publish("voice:keyword_detected", event);

    // Auto-switch to active mode
    m_sttService->setMode(STTMode::ACTIVE);
}

void VoiceService::handleTranscription(const std::string& text, STTMode mode) {
    m_logger->info("Transcription ({}): {}",
        mode == STTMode::PASSIVE ? "passive" : "active", text);

    // Publish transcription
    nlohmann::json event = {
        {"type", "transcription"},
        {"text", text},
        {"mode", mode == STTMode::PASSIVE ? "passive" : "active"},
        {"timestamp", std::time(nullptr)}
    };
    m_io->publish("voice:transcription", event);
}
```

**Estimation modifications** : +150 lignes

---

## Phase 7.4 - Configuration

### Fichier à modifier

#### `config/voice.json`

**Configuration complète** :

```json
{
  "tts": {
    "enabled": true,
    "engine": "auto",
    "rate": 0,
    "volume": 80,
    "voice": "fr-fr"
  },
  "stt": {
    "passive_mode": {
      "enabled": true,
      "engine": "pocketsphinx",
      "keywords": ["celuna", "hey celuna", "ok celuna"],
      "threshold": 0.8,
      "model_path": "/usr/share/pocketsphinx/model/en-us"
    },
    "active_mode": {
      "enabled": true,
      "engine": "vosk",
      "model_path": "./models/vosk-model-small-fr-0.22",
      "language": "fr",
      "timeout_seconds": 30,
      "fallback_engine": "whisper-api"
    },
    "whisper_api": {
      "api_key_env": "OPENAI_API_KEY",
      "model": "whisper-1"
    },
    "microphone": {
      "device_id": -1,
      "sample_rate": 16000,
      "channels": 1,
      "buffer_size": 1024
    }
  }
}
```

---

## Phase 7.5 - Tests

### Fichiers à créer

#### `tests/services/STTServiceTests.cpp`

**Tests unitaires** :
- ✅ Création service
- ✅ Start/stop
- ✅ Switch passive/active
- ✅ Keyword detection
- ✅ Transcription
- ✅ Fallback engine
- ✅ Timeout active → passive

**Estimation** : 200 lignes

---

#### `tests/integration/IT_014_VoicePassiveMode.cpp`

**Test d'intégration passive mode** :

```cpp
// Simulate audio avec keyword "celuna"
// Vérifie :
//   1. PocketSphinx détecte keyword
//   2. Event "voice:keyword_detected" publié
//   3. Switch vers ACTIVE mode
//   4. Timeout 30s → retour PASSIVE
```

**Estimation** : 150 lignes

---

#### `tests/integration/IT_015_VoiceActiveTranscription.cpp`

**Test d'intégration active mode** :

```cpp
// Simulate conversation complète :
//   1. User: "celuna" → keyword detected
//   2. User: "quelle heure est-il ?" → transcription via Vosk
//   3. AI responds → TTS
//   4. Timeout → retour passive
```

**Estimation** : 200 lignes

---

## Phase 7.6 - Documentation

### Fichiers à créer/modifier

#### `docs/STT_ARCHITECTURE.md`

**Documentation technique** :
- Architecture STT
- Choix engines
- Configuration
- Troubleshooting

**Estimation** : 400 lignes

---

#### `README.md`

**Mise à jour roadmap** :

```markdown
### Completed ✅
- [x] STT multi-engine (Vosk, PocketSphinx, Whisper)
- [x] Passive/Active mode (keyword "Celuna")
- [x] Local STT (coût zéro)
```

---

## Récapitulatif Estimation

| Tâche | Fichiers | Lignes | Priorité |
|-------|----------|--------|----------|
| **7.1 Service Layer** | `ISTTService.hpp`, `STTService.{h,cpp}` | 350 | P0 |
| **7.2 Vosk Engine** | `VoskSTTEngine.hpp` | 200 | P0 |
| **7.2 PocketSphinx** | `PocketSphinxEngine.hpp` | 180 | P1 |
| **7.2 WhisperCpp** | `WhisperCppEngine.hpp` | 250 | P2 (optionnel) |
| **7.2 Factory** | `STTEngineFactory.cpp` | 80 | P0 |
| **7.3 VoiceService** | `VoiceService.cpp` (modifs) | +150 | P0 |
| **7.4 Config** | `voice.json` | +30 | P0 |
| **7.5 Tests unitaires** | `STTServiceTests.cpp` | 200 | P1 |
| **7.5 Tests intégration** | `IT_014`, `IT_015` | 350 | P1 |
| **7.6 Documentation** | `STT_ARCHITECTURE.md`, README | 450 | P2 |
| **TOTAL** | 14 fichiers | **~2240 lignes** | |

---

## Plan d'Exécution

### Milestone 1 : MVP STT Local (Vosk seul) ⚡

**Objectif** : STT fonctionnel sans keyword detection

**Tâches** :
1. ✅ Créer `ISTTService.hpp`
2. ✅ Créer `STTService` (simple, sans passive mode)
3. ✅ Créer `VoskSTTEngine`
4. ✅ Modifier `STTEngineFactory`
5. ✅ Intégrer dans `VoiceService`
6. ✅ Config `voice.json`
7. ✅ Test manuel transcription

**Durée estimée** : 3-4h
**Lignes** : ~600

---

### Milestone 2 : Passive Mode (Keyword Detection) 🎧

**Objectif** : Détection "Celuna" + switch auto

**Tâches** :
1. ✅ Créer `PocketSphinxEngine`
2. ✅ Étendre `STTService` (dual mode)
3. ✅ Callbacks keyword/transcription
4. ✅ Timeout active → passive
5. ✅ Config passive/active
6. ✅ Tests IT_014, IT_015

**Durée estimée** : 4-5h
**Lignes** : ~700

---

### Milestone 3 : Fallback Whisper API 🔄

**Objectif** : Robustesse avec fallback cloud

**Tâches** :
1. ✅ Intégrer `WhisperAPIEngine` existant
2. ✅ Logique fallback dans `STTService`
3. ✅ Config fallback
4. ✅ Tests fallback

**Durée estimée** : 2h
**Lignes** : ~200

---

### Milestone 4 : Polish & Documentation 📝

**Tâches** :
1. ✅ Documentation complète
2. ✅ Tests unitaires STTService
3. ✅ Troubleshooting guide
4. ✅ Mise à jour README

**Durée estimée** : 3h
**Lignes** : ~700

---

## Dépendances Externes

### À installer

```bash
# Vosk
sudo apt install libvosk-dev
wget https://alphacephei.com/vosk/models/vosk-model-small-fr-0.22.zip
unzip vosk-model-small-fr-0.22.zip -d models/

# PocketSphinx
sudo apt install pocketsphinx pocketsphinx-en-us

# PortAudio (pour microphone)
sudo apt install portaudio19-dev

# Optionnel: whisper.cpp
git clone https://github.com/ggerganov/whisper.cpp external/whisper.cpp
cd external/whisper.cpp && make
```

### CMakeLists.txt

```cmake
# Find Vosk
find_library(VOSK_LIBRARY vosk REQUIRED)
find_path(VOSK_INCLUDE_DIR vosk_api.h REQUIRED)

# Find PocketSphinx
find_library(POCKETSPHINX_LIBRARY pocketsphinx REQUIRED)
find_path(POCKETSPHINX_INCLUDE_DIR pocketsphinx.h REQUIRED)

# Find PortAudio
find_library(PORTAUDIO_LIBRARY portaudio REQUIRED)
find_path(PORTAUDIO_INCLUDE_DIR portaudio.h REQUIRED)

# Link
target_link_libraries(VoiceService
    ${VOSK_LIBRARY}
    ${POCKETSPHINX_LIBRARY}
    ${PORTAUDIO_LIBRARY}
)
```

---

## Risques & Mitigation

| Risque | Impact | Mitigation |
|--------|--------|------------|
| **Vosk model trop lourd** | RAM (50MB) | Utiliser `vosk-model-small` au lieu de `base` |
| **PocketSphinx faux positifs** | UX | Ajuster threshold (1e-40 → 1e-50) |
| **Microphone permissions** | Bloquant | Guide installation PortAudio + permissions |
| **Latence transcription** | UX | Buffer 1-2s audio avant transcription |
| **Whisper API coût** | Budget | Utiliser seulement en fallback (rare) |

---

## Prochaines Étapes

**Après validation de ce plan** :

1. **Installer dépendances** (Vosk, PocketSphinx, PortAudio)
2. **Milestone 1** : Vosk STT basique
3. **Tester** : Transcription fichier audio FR
4. **Milestone 2** : Keyword "Celuna"
5. **Tester** : Conversation complète passive → active
6. **Commit + Push** : Phase 7 complète

---

## Validation Plan

**Questions avant implémentation** :

1. ✅ Architecture service layer approuvée ?
2. ✅ Choix engines (Vosk + PocketSphinx) OK ?
3. ❓ Besoin WhisperCpp ou Vosk suffit ?
4. ✅ Nom "Celuna" confirmé ?
5. ❓ Autres keywords à détecter ("hey celuna", "ok celuna") ?

---

**Auteur** : Claude Code
**Date** : 2025-11-29
**Phase** : 7 - STT Implementation
**Status** : 📋 Plan - En attente validation