ShieldAI/apps/api/src/services/voiceprint/voiceprint.service.ts

import { prisma, VoiceEnrollment, VoiceAnalysis } from '@shieldsai/shared-db';
import {
  voicePrintEnv,
  AnalysisJobStatus,
  DetectionType,
  ConfidenceLevel,
  audioPreprocessingConfig,
  voicePrintFeatureFlags,
} from './voiceprint.config';
import { checkFlag } from './voiceprint.feature-flags';

// Audio preprocessing service
export class AudioPreprocessor {
  /**
   * Normalize audio to 16kHz mono with VAD and noise reduction.
   * Returns preprocessing metadata and the processed audio buffer.
   */
  async preprocess(
    audioBuffer: Buffer,
    options?: {
      sourceSampleRate?: number;
      channels?: number;
    }
  ): Promise<{
    buffer: Buffer;
    metadata: {
      sampleRate: number;
      channels: number;
      duration: number;
      format: string;
    };
  }> {
    const duration = this.estimateDuration(audioBuffer, options?.sourceSampleRate ?? 44100);

    if (duration < voicePrintEnv.ENROLLMENT_MIN_DURATION_SEC) {
      throw new Error(
        `Audio too short: ${duration.toFixed(1)}s < ${voicePrintEnv.ENROLLMENT_MIN_DURATION_SEC}s minimum`
      );
    }

    if (duration > voicePrintEnv.ENROLLMENT_MAX_DURATION_SEC) {
      throw new Error(
        `Audio too long: ${duration.toFixed(1)}s > ${voicePrintEnv.ENROLLMENT_MAX_DURATION_SEC}s maximum`
      );
    }

    // TODO: Integrate with Python librosa/torchaudio for actual preprocessing
    // For MVP, return original buffer with target metadata
    return {
      buffer: audioBuffer,
      metadata: {
        sampleRate: audioPreprocessingConfig.sampleRate,
        channels: audioPreprocessingConfig.channels,
        duration,
        format: 'wav',
      },
    };
  }

  /**
   * Apply Voice Activity Detection to remove silence segments.
   */
  async applyVAD(buffer: Buffer): Promise<Buffer> {
    // TODO: Integrate with Python webrtcvad or silero-vad
    // For MVP, return original buffer
    return buffer;
  }

  /**
   * Estimate audio duration from buffer size and sample rate.
   */
  private estimateDuration(
    buffer: Buffer,
    sampleRate: number
  ): number {
    const bytesPerSample = 2;
    const channels = 1;
    const samples = buffer.length / (bytesPerSample * channels);
    return samples / sampleRate;
  }
}

// Voice enrollment service
export class VoiceEnrollmentService {
  /**
   * Enroll a new voice profile from audio data.
   */
  async enroll(
    userId: string,
    name: string,
    audioBuffer: Buffer
  ): Promise<VoiceEnrollment> {
    const preprocessor = new AudioPreprocessor();
    const processed = await preprocessor.preprocess(audioBuffer);

    const embeddingService = new EmbeddingService();
    const embedding = await embeddingService.extract(processed.buffer);
    const voiceHash = this.computeEmbeddingHash(embedding);

    const enrollment = await prisma.voiceEnrollment.create({
      data: {
        userId,
        name,
        voiceHash,
        audioMetadata: {
          ...processed.metadata,
          embeddingDimensions: embedding.length,
          enrollmentTimestamp: new Date().toISOString(),
        },
      },
    });

    // Index in FAISS for similarity search
    const faissIndex = new FAISSIndex();
    await faissIndex.add(enrollment.id, embedding);

    return enrollment;
  }

  /**
   * List all enrollments for a user.
   */
  async listEnrollments(
    userId: string,
    options?: {
      isActive?: boolean;
      limit?: number;
      offset?: number;
    }
  ): Promise<VoiceEnrollment[]> {
    return prisma.voiceEnrollment.findMany({
      where: {
        userId,
        ...(options?.isActive !== undefined && { isActive: options.isActive }),
      },
      orderBy: { createdAt: 'desc' },
      take: options?.limit ?? 50,
      skip: options?.offset ?? 0,
    });
  }

  /**
   * Get a single enrollment by ID.
   */
  async getEnrollment(
    enrollmentId: string,
    userId: string
  ): Promise<VoiceEnrollment | null> {
    return prisma.voiceEnrollment.findFirst({
      where: {
        id: enrollmentId,
        userId,
      },
    });
  }

  /**
   * Remove (deactivate) an enrollment.
   */
  async removeEnrollment(
    enrollmentId: string,
    userId: string
  ): Promise<VoiceEnrollment> {
    const enrollment = await this.getEnrollment(enrollmentId, userId);
    if (!enrollment) {
      throw new Error('Enrollment not found');
    }

    const faissIndex = new FAISSIndex();
    await faissIndex.remove(enrollmentId);

    return prisma.voiceEnrollment.update({
      where: { id: enrollmentId },
      data: { isActive: false },
    });
  }

  /**
   * Search for similar enrollments using FAISS.
   */
  async findSimilar(
    embedding: number[],
    topK: number = 5
  ): Promise<Array<{ enrollment: VoiceEnrollment; similarity: number }>> {
    const faissIndex = new FAISSIndex();
    const results = await faissIndex.search(embedding, topK);

    const enrollmentIds = results.map((r) => r.id);
    const enrollments = await prisma.voiceEnrollment.findMany({
      where: { id: { in: enrollmentIds } },
    });

    return results.map((r, i) => ({
      enrollment: enrollments[i],
      similarity: r.similarity,
    }));
  }

  private computeEmbeddingHash(embedding: number[]): string {
    let hash = 0;
    for (let i = 0; i < embedding.length; i++) {
      hash = ((hash << 5) - hash) + embedding[i];
      hash |= 0;
    }
    return `vp_${Math.abs(hash).toString(16)}_${embedding.length}`;
  }
}

// Audio analysis service
export class AnalysisService {
  /**
   * Analyze a single audio file for synthetic voice detection.
   */
  async analyze(
    userId: string,
    audioBuffer: Buffer,
    options?: {
      enrollmentId?: string;
      audioUrl?: string;
    }
  ): Promise<VoiceAnalysis> {
    const preprocessor = new AudioPreprocessor();
    const processed = await preprocessor.preprocess(audioBuffer);

    const audioHash = this.computeAudioHash(audioBuffer);

    const embeddingService = new EmbeddingService();
    const analysisResult = await embeddingService.analyze(processed.buffer);

    const isSynthetic = analysisResult.confidence >= voicePrintEnv.SYNTHETIC_THRESHOLD;

    const voiceAnalysis = await prisma.voiceAnalysis.create({
      data: {
        userId,
        enrollmentId: options?.enrollmentId,
        audioHash,
        isSynthetic,
        confidence: analysisResult.confidence,
        analysisResult: {
          ...analysisResult,
          processedMetadata: processed.metadata,
          analysisTimestamp: new Date().toISOString(),
          modelVersion: 'ecapa-tdnn-v1-mock',
        },
        audioUrl: options?.audioUrl ?? '',
      },
    });

    return voiceAnalysis;
  }

  /**
   * Get analysis result by ID.
   */
  async getResult(
    analysisId: string,
    userId: string
  ): Promise<VoiceAnalysis | null> {
    return prisma.voiceAnalysis.findFirst({
      where: {
        id: analysisId,
        userId,
      },
    });
  }

  /**
   * Get analysis history for a user.
   */
  async getHistory(
    userId: string,
    options?: {
      limit?: number;
      offset?: number;
      isSynthetic?: boolean;
    }
  ): Promise<VoiceAnalysis[]> {
    return prisma.voiceAnalysis.findMany({
      where: {
        userId,
        ...(options?.isSynthetic !== undefined && { isSynthetic: options.isSynthetic }),
      },
      orderBy: { createdAt: 'desc' },
      take: options?.limit ?? 50,
      skip: options?.offset ?? 0,
    });
  }

  private computeAudioHash(buffer: Buffer): string {
    let hash = 0;
    const sampleSize = Math.min(buffer.length, 1024);
    for (let i = 0; i < sampleSize; i += 8) {
      hash = ((hash << 5) - hash) + buffer.readUInt8(i);
      hash |= 0;
    }
    return `audio_${Math.abs(hash).toString(16)}`;
  }
}

// Batch analysis service
export class BatchAnalysisService {
  /**
   * Analyze multiple audio files in a batch.
   */
  async analyzeBatch(
    userId: string,
    files: Array<{
      name: string;
      buffer: Buffer;
      audioUrl?: string;
    }>,
    options?: {
      enrollmentId?: string;
    }
  ): Promise<{
    jobId: string;
    results: VoiceAnalysis[];
    summary: {
      total: number;
      synthetic: number;
      natural: number;
      failed: number;
    };
  }> {
    if (files.length > voicePrintEnv.BATCH_MAX_FILES) {
      throw new Error(
        `Batch too large: ${files.length} > ${voicePrintEnv.BATCH_MAX_FILES} max`
      );
    }

    const analysisService = new AnalysisService();
    const results: VoiceAnalysis[] = [];
    let synthetic = 0;
    let natural = 0;
    let failed = 0;

    for (const file of files) {
      try {
        const result = await analysisService.analyze(userId, file.buffer, {
          enrollmentId: options?.enrollmentId,
          audioUrl: file.audioUrl,
        });
        results.push(result);
        if (result.isSynthetic) {
          synthetic++;
        } else {
          natural++;
        }
      } catch (error) {
        console.error(`Batch analysis failed for ${file.name}:`, error);
        failed++;
      }
    }

    const jobId = `batch_${Date.now()}_${Math.random().toString(36).slice(2, 8)}`;

    return {
      jobId,
      results,
      summary: {
        total: files.length,
        synthetic,
        natural,
        failed,
      },
    };
  }
}

// Embedding service — ECAPA-TDNN inference wrapper
export class EmbeddingService {
  private initialized = false;

  /**
   * Initialize the ECAPA-TDNN model.
   */
  async initialize(): Promise<void> {
    if (this.initialized) return;

    // TODO: Connect to Python ML service for real inference
    // const response = await fetch(`${voicePrintEnv.ML_SERVICE_URL}/initialize`, {
    //   method: 'POST',
    //   body: JSON.stringify({ modelPath: voicePrintEnv.ECAPA_TDNN_MODEL_PATH }),
    // });

    this.initialized = true;
    console.log('Embedding service initialized (mock model)');
  }

  /**
   * Extract voice embedding from audio.
   */
  async extract(audioBuffer: Buffer): Promise<number[]> {
    await this.initialize();

    // TODO: Call Python ML service
    // const response = await fetch(`${voicePrintEnv.ML_SERVICE_URL}/embed`, {
    //   method: 'POST',
    //   body: audioBuffer,
    // });
    // const data = await response.json();
    // return data.embedding;

    // Mock: generate deterministic embedding based on buffer content
    const dims = voicePrintEnv.EMBEDDING_DIMENSIONS;
    const embedding: number[] = new Array(dims);
    let hash = 0;
    for (let i = 0; i < Math.min(audioBuffer.length, 256); i++) {
      hash = ((hash << 5) - hash) + audioBuffer[i];
      hash |= 0;
    }
    for (let i = 0; i < dims; i++) {
      hash = ((hash << 5) - hash) + i;
      hash |= 0;
      embedding[i] = (Math.abs(hash) % 1000) / 1000.0;
    }

    // L2 normalize
    const norm = Math.sqrt(embedding.reduce((s, v) => s + v * v, 0));
    return embedding.map((v) => v / norm);
  }

  /**
   * Run full analysis: embedding + synthetic detection.
   */
  async analyze(audioBuffer: Buffer): Promise<{
    confidence: number;
    detectionType: DetectionType;
    features: Record<string, number>;
    embedding: number[];
  }> {
    const embedding = await this.extract(audioBuffer);

    // TODO: Run synthetic voice detection model
    // For MVP, use heuristic based on embedding statistics
    const confidence = this.estimateSyntheticConfidence(audioBuffer, embedding);
    const detectionType =
      confidence >= voicePrintEnv.SYNTHETIC_THRESHOLD
        ? DetectionType.SYNTHETIC_VOICE
        : DetectionType.NATURAL;

    const features = this.extractAnalysisFeatures(audioBuffer, embedding);

    return {
      confidence,
      detectionType,
      features,
      embedding,
    };
  }

  private estimateSyntheticConfidence(
    buffer: Buffer,
    embedding: number[]
  ): number {
    // Heuristic features for synthetic detection
    const meanAmplitude =
      buffer.reduce((s, v) => s + v, 0) / buffer.length / 255;
    const embeddingStdDev =
      Math.sqrt(
        embedding.reduce((s, v) => s + (v - embedding.reduce((a, b) => a + b) / embedding.length) ** 2, 0) /
          embedding.length
      ) || 0;

    // Combine features into confidence score
    const amplitudeScore = Math.abs(meanAmplitude - 0.5) * 2;
    const embeddingScore = 1.0 - Math.min(1.0, embeddingStdDev * 2);

    return Math.min(
      1.0,
      amplitudeScore * 0.3 + embeddingScore * 0.4 + Math.random() * 0.3
    );
  }

  private extractAnalysisFeatures(
    buffer: Buffer,
    embedding: number[]
  ): Record<string, number> {
    const meanAmplitude =
      buffer.reduce((s, v) => s + v, 0) / buffer.length / 255;
    const zeroCrossings = buffer.reduce((count, v, i, arr) => {
      return i > 0 && ((v - 128) * (arr[i - 1] - 128) < 0) ? count + 1 : count;
    }, 0);

    return {
      mean_amplitude: meanAmplitude,
      zero_crossing_rate: zeroCrossings / buffer.length,
      embedding_energy: embedding.reduce((s, v) => s + v * v, 0),
      embedding_entropy: this.calculateEntropy(embedding),
    };
  }

  private calculateEntropy(values: number[]): number {
    const bins = 20;
    const histogram = new Array(bins).fill(0);
    const min = Math.min(...values);
    const max = Math.max(...values);
    const range = max - min || 1;

    for (const v of values) {
      const bin = Math.min(bins - 1, Math.floor(((v - min) / range) * bins));
      histogram[bin]++;
    }

    let entropy = 0;
    const total = values.length;
    for (const count of histogram) {
      if (count > 0) {
        const p = count / total;
        entropy -= p * Math.log2(p);
      }
    }
    return entropy;
  }
}

// FAISS index wrapper for voice fingerprint matching
export class FAISSIndex {
  private indexPath: string;
  private initialized = false;

  constructor(path?: string) {
    this.indexPath = path ?? voicePrintEnv.FAISS_INDEX_PATH;
  }

  /**
   * Initialize or load the FAISS index.
   */
  async initialize(): Promise<void> {
    if (this.initialized) return;

    // TODO: Load FAISS index from disk
    // const faiss = require('faiss-node');
    // this.index = faiss.readIndex(this.indexPath);

    this.initialized = true;
    console.log(`FAISS index initialized at ${this.indexPath}`);
  }

  /**
   * Add an enrollment embedding to the index.
   */
  async add(enrollmentId: string, embedding: number[]): Promise<void> {
    await this.initialize();

    // TODO: Add to FAISS index
    // this.index.add([embedding]);
    // Store mapping: enrollmentId -> index position
    console.log(`Added enrollment ${enrollmentId} to FAISS index`);
  }

  /**
   * Remove an enrollment from the index.
   */
  async remove(enrollmentId: string): Promise<void> {
    await this.initialize();

    // TODO: Remove from FAISS index
    console.log(`Removed enrollment ${enrollmentId} from FAISS index`);
  }

  /**
   * Search for similar voice embeddings.
   */
  async search(
    embedding: number[],
    topK: number = 5
  ): Promise<Array<{ id: string; similarity: number }>> {
    await this.initialize();

    // TODO: Query FAISS index
    // const [distances, indices] = this.index.search([embedding], topK);
    // Map indices back to enrollment IDs

    // Mock: return empty results
    return [];
  }

  /**
   * Save the index to disk.
   */
  async save(): Promise<void> {
    await this.initialize();
    // TODO: Write FAISS index to disk
    console.log(`FAISS index saved to ${this.indexPath}`);
  }
}

// Export singleton instances
export const audioPreprocessor = new AudioPreprocessor();
export const voiceEnrollmentService = new VoiceEnrollmentService();
export const analysisService = new AnalysisService();
export const batchAnalysisService = new BatchAnalysisService();
export const embeddingService = new EmbeddingService();