Deepgram パフォーマンス チューニング

概要

オーディオの前処理、接続管理、設定チューニングを通じて Deepgram 統合のパフォーマンスを最適化します。

前提条件

• 動作する Deepgram 統合
• パフォーマンス監視の実施
• オーディオ処理機能
• ベースラインメトリクスの確立

パフォーマンス要因

要因	影響	最適化
オーディオ形式	高	最適なエンコーディングを使用
サンプリングレート	中	モデル要件に合わせる
ファイルサイズ	高	大きなファイルはストリーミング
モデル選択	高	精度と速度のバランスを調整
ネットワークレイテンシ	中	最も近いリージョンを使用
同時実行数	中	接続を管理

手順

ステップ 1: オーディオ形式を最適化

最適な文字起こしのためにオーディオを前処理します。

ステップ 2: 接続プーリングを設定

接続を再利用してスループットを向上させます。

ステップ 3: API パラメータをチューニング

適切なモデルと機能を選択します。

ステップ 4: ストリーミングを実装

リアルタイムと大きなファイル用にストリーミングを使用します。

例

オーディオ前処理

// lib/audio-optimizer.ts
import ffmpeg from 'fluent-ffmpeg';
import { Readable } from 'stream';

interface OptimizedAudio {
  buffer: Buffer;
  mimetype: string;
  sampleRate: number;
  channels: number;
  duration: number;
}

export async function optimizeAudio(inputPath: string): Promise<OptimizedAudio> {
  return new Promise((resolve, reject) => {
    const chunks: Buffer[] = [];

    // Optimal settings for Deepgram
    ffmpeg(inputPath)
      .audioCodec('pcm_s16le')      // 16-bit PCM
      .audioChannels(1)              // Mono
      .audioFrequency(16000)         // 16kHz (optimal for speech)
      .format('wav')
      .on('error', reject)
      .on('end', () => {
        const buffer = Buffer.concat(chunks);
        resolve({
          buffer,
          mimetype: 'audio/wav',
          sampleRate: 16000,
          channels: 1,
          duration: buffer.length / (16000 * 2), // 16-bit = 2 bytes
        });
      })
      .pipe()
      .on('data', (chunk: Buffer) => chunks.push(chunk));
  });
}

// For already loaded audio data
export async function optimizeAudioBuffer(
  audioBuffer: Buffer,
  inputFormat: string
): Promise<Buffer> {
  return new Promise((resolve, reject) => {
    const chunks: Buffer[] = [];
    const readable = new Readable();
    readable.push(audioBuffer);
    readable.push(null);

    ffmpeg(readable)
      .inputFormat(inputFormat)
      .audioCodec('pcm_s16le')
      .audioChannels(1)
      .audioFrequency(16000)
      .format('wav')
      .on('error', reject)
      .on('end', () => resolve(Buffer.concat(chunks)))
      .pipe()
      .on('data', (chunk: Buffer) => chunks.push(chunk));
  });
}

接続プーリング

// lib/connection-pool.ts
import { createClient, DeepgramClient } from '@deepgram/sdk';

interface PoolConfig {
  minSize: number;
  maxSize: number;
  acquireTimeout: number;
  idleTimeout: number;
}

class DeepgramConnectionPool {
  private pool: DeepgramClient[] = [];
  private inUse: Set<DeepgramClient> = new Set();
  private waiting: Array<(client: DeepgramClient) => void> = [];
  private config: PoolConfig;
  private apiKey: string;

  constructor(apiKey: string, config: Partial<PoolConfig> = {}) {
    this.apiKey = apiKey;
    this.config = {
      minSize: config.minSize ?? 2,
      maxSize: config.maxSize ?? 10,
      acquireTimeout: config.acquireTimeout ?? 10000,
      idleTimeout: config.idleTimeout ?? 60000,
    };

    // Initialize minimum connections
    for (let i = 0; i < this.config.minSize; i++) {
      this.pool.push(createClient(this.apiKey));
    }
  }

  async acquire(): Promise<DeepgramClient> {
    // Try to get from pool
    if (this.pool.length > 0) {
      const client = this.pool.pop()!;
      this.inUse.add(client);
      return client;
    }

    // Create new if under max
    if (this.inUse.size < this.config.maxSize) {
      const client = createClient(this.apiKey);
      this.inUse.add(client);
      return client;
    }

    // Wait for available connection
    return new Promise((resolve, reject) => {
      const timeout = setTimeout(() => {
        const index = this.waiting.indexOf(resolve);
        if (index > -1) this.waiting.splice(index, 1);
        reject(new Error('Connection acquire timeout'));
      }, this.config.acquireTimeout);

      this.waiting.push((client) => {
        clearTimeout(timeout);
        resolve(client);
      });
    });
  }

  release(client: DeepgramClient): void {
    this.inUse.delete(client);

    if (this.waiting.length > 0) {
      const waiter = this.waiting.shift()!;
      this.inUse.add(client);
      waiter(client);
    } else {
      this.pool.push(client);
    }
  }

  async execute<T>(fn: (client: DeepgramClient) => Promise<T>): Promise<T> {
    const client = await this.acquire();
    try {
      return await fn(client);
    } finally {
      this.release(client);
    }
  }

  getStats() {
    return {
      poolSize: this.pool.length,
      inUse: this.inUse.size,
      waiting: this.waiting.length,
    };
  }
}

export const pool = new DeepgramConnectionPool(process.env.DEEPGRAM_API_KEY!);

大きなファイル用ストリーミング

// lib/streaming-transcription.ts
import { createClient } from '@deepgram/sdk';
import { createReadStream, statSync } from 'fs';

interface StreamingOptions {
  chunkSize: number;
  model: string;
}

export async function streamLargeFile(
  filePath: string,
  options: Partial<StreamingOptions> = {}
): Promise<string> {
  const { chunkSize = 1024 * 1024, model = 'nova-2' } = options;
  const client = createClient(process.env.DEEPGRAM_API_KEY!);

  const fileSize = statSync(filePath).size;
  const transcripts: string[] = [];

  // Use live transcription for streaming
  const connection = client.listen.live({
    model,
    smart_format: true,
    punctuate: true,
  });

  return new Promise((resolve, reject) => {
    connection.on('open', () => {
      const stream = createReadStream(filePath, { highWaterMark: chunkSize });

      stream.on('data', (chunk: Buffer) => {
        connection.send(chunk);
      });

      stream.on('end', () => {
        connection.finish();
      });

      stream.on('error', reject);
    });

    connection.on('transcript', (data) => {
      if (data.is_final) {
        transcripts.push(data.channel.alternatives[0].transcript);
      }
    });

    connection.on('close', () => {
      resolve(transcripts.join(' '));
    });

    connection.on('error', reject);
  });
}

速度のためのモデル選択

// lib/model-selector.ts
interface ModelConfig {
  name: string;
  accuracy: 'high' | 'medium' | 'low';
  speed: 'fast' | 'medium' | 'slow';
  costPerMinute: number;
}

const models: Record<string, ModelConfig> = {
  'nova-2': {
    name: 'Nova-2',
    accuracy: 'high',
    speed: 'fast',
    costPerMinute: 0.0043,
  },
  'nova': {
    name: 'Nova',
    accuracy: 'high',
    speed: 'fast',
    costPerMinute: 0.0043,
  },
  'enhanced': {
    name: 'Enhanced',
    accuracy: 'medium',
    speed: 'fast',
    costPerMinute: 0.0145,
  },
  'base': {
    name: 'Base',
    accuracy: 'low',
    speed: 'fast',
    costPerMinute: 0.0048,
  },
};

export function selectModel(requirements: {
  prioritize: 'accuracy' | 'speed' | 'cost';
  minAccuracy?: 'high' | 'medium' | 'low';
}): string {
  const { prioritize, minAccuracy = 'low' } = requirements;

  const accuracyOrder = ['high', 'medium', 'low'];
  const minAccuracyIndex = accuracyOrder.indexOf(minAccuracy);

  const eligible = Object.entries(models).filter(([_, config]) =>
    accuracyOrder.indexOf(config.accuracy) <= minAccuracyIndex
  );

  if (prioritize === 'accuracy') {
    return eligible.reduce((best, [name, config]) =>
      accuracyOrder.indexOf(config.accuracy) < accuracyOrder.indexOf(models[best].accuracy)
        ? name : best
    , eligible[0][0]);
  }

  if (prioritize === 'cost') {
    return eligible.reduce((best, [name, config]) =>
      config.costPerMinute < models[best].costPerMinute ? name : best
    , eligible[0][0]);
  }

  // Default: balance speed and accuracy
  return 'nova-2';
}

並列処理

// lib/parallel-transcription.ts
import { pool } from './connection-pool';
import pLimit from 'p-limit';

interface TranscriptionResult {
  file: string;
  transcript: string;
  duration: number;
}

export async function transcribeMultiple(
  audioUrls: string[],
  concurrency = 5
): Promise<TranscriptionResult[]> {
  const limit = pLimit(concurrency);
  const startTime = Date.now();

  const results = await Promise.all(
    audioUrls.map((url, index) =>
      limit(async () => {
        const itemStart = Date.now();

        const result = await pool.execute(async (client) => {
          const { result, error } = await client.listen.prerecorded.transcribeUrl(
            { url },
            { model: 'nova-2', smart_format: true }
          );

          if (error) throw error;
          return result;
        });

        return {
          file: url,
          transcript: result.results.channels[0].alternatives[0].transcript,
          duration: Date.now() - itemStart,
        };
      })
    )
  );

  console.log(`Processed ${audioUrls.length} files in ${Date.now() - startTime}ms`);
  console.log(`Average per file: ${(Date.now() - startTime) / audioUrls.length}ms`);

  return results;
}

結果のキャッシング

// lib/transcription-cache.ts
import { createHash } from 'crypto';
import { redis } from './redis';

interface CacheOptions {
  ttl: number; // seconds
}

export class TranscriptionCache {
  private ttl: number;

  constructor(options: Partial<CacheOptions> = {}) {
    this.ttl = options.ttl ?? 3600; // 1 hour default
  }

  private getCacheKey(audioUrl: string, options: Record<string, unknown>): string {
    const hash = createHash('sha256')
      .update(JSON.stringify({ audioUrl, options }))
      .digest('hex');
    return `transcription:${hash}`;
  }

  async get(
    audioUrl: string,
    options: Record<string, unknown>
  ): Promise<string | null> {
    const key = this.getCacheKey(audioUrl, options);
    return redis.get(key);
  }

  async set(
    audioUrl: string,
    options: Record<string, unknown>,
    transcript: string
  ): Promise<void> {
    const key = this.getCacheKey(audioUrl, options);
    await redis.setex(key, this.ttl, transcript);
  }

  async transcribeWithCache(
    transcribeFn: () => Promise<string>,
    audioUrl: string,
    options: Record<string, unknown>
  ): Promise<{ transcript: string; cached: boolean }> {
    const cached = await this.get(audioUrl, options);
    if (cached) {
      return { transcript: cached, cached: true };
    }

    const transcript = await transcribeFn();
    await this.set(audioUrl, options, transcript);

    return { transcript, cached: false };
  }
}

パフォーマンスメトリクス

// lib/performance-metrics.ts
import { Histogram, Counter, Gauge } from 'prom-client';

export const transcriptionLatency = new Histogram({
  name: 'deepgram_transcription_latency_seconds',
  help: 'Latency of transcription requests',
  labelNames: ['model', 'status'],
  buckets: [0.5, 1, 2, 5, 10, 30, 60],
});

export const audioDuration = new Histogram({
  name: 'deepgram_audio_duration_seconds',
  help: 'Duration of audio files processed',
  buckets: [10, 30, 60, 120, 300, 600, 1800],
});

export const processingRatio = new Gauge({
  name: 'deepgram_processing_ratio',
  help: 'Ratio of processing time to audio duration',
  labelNames: ['model'],
});

export function measureTranscription(
  audioDurationSec: number,
  processingTimeSec: number,
  model: string
) {
  audioDuration.observe(audioDurationSec);
  processingRatio.labels(model).set(processingTimeSec / audioDurationSec);
}

リソース

• Deepgram パフォーマンスガイド
• オーディオ形式のベストプラクティス
• FFmpeg ドキュメント

次のステップ

コスト最適化については deepgram-cost-tuning に進みます。