pipeline/packages/tts/src/providers/openai-tts.ts

import type {
  TTSSynthesizeRequest,
  TTSResult,
  VoiceInfo,
  WordTimestamp,
} from "@pipeline/shared";
import type { TTSProvider } from "../types.js";
import { registerProvider } from "../registry.js";
import { measureAudioDuration } from "../duration.js";
import { fetchWithRetry } from "../retry.js";

class OpenAITTSProvider implements TTSProvider {
  readonly name = "openai-tts";

  async synthesize(request: TTSSynthesizeRequest): Promise<TTSResult> {
    const apiKey = process.env.OPENAI_TTS_API_KEY;
    if (!apiKey) throw new Error("OPENAI_TTS_API_KEY not set");

    const baseURL =
      process.env.OPENAI_TTS_BASE_URL ||
      "https://new-api.corp.shuidi.tech/v1";
    const model = process.env.OPENAI_TTS_MODEL || "seed-tts-1.1";

    const response = await fetchWithRetry(`${baseURL}/audio/speech`, {
      method: "POST",
      headers: {
        "Content-Type": "application/json",
        Authorization: `Bearer ${apiKey}`,
      },
      body: JSON.stringify({
        input: request.text,
        model,
        response_format: request.format || "mp3",
        speed: request.speed ?? 1.0,
        voice: request.voiceId || "zh_female_linjianvhai_uranus_bigtts",
      }),
    });

    if (!response.ok) {
      const body = await response.text();
      throw new Error(`OpenAI TTS API error: ${response.status} ${body}`);
    }

    const audioBuffer = Buffer.from(await response.arrayBuffer());
    const format = request.format || "mp3";
    const outputPath = `${request.outputDir}/${request.filename}.${format}`;

    const { writeFile, mkdir } = await import("node:fs/promises");
    await mkdir(request.outputDir, { recursive: true });
    await writeFile(outputPath, audioBuffer);

    const durationSeconds = await measureAudioDuration(outputPath);

    return {
      audioFilePath: outputPath,
      durationSeconds,
      wordTimestamps: estimateWordTimestamps(request.text, durationSeconds),
      format,
    };
  }

  async listVoices(): Promise<VoiceInfo[]> {
    // Voice IDs are configured via Volcengine docs:
    // https://www.volcengine.com/docs/6561/1257544
    return [];
  }

  async align(audio: Buffer, text: string): Promise<WordTimestamp[]> {
    return estimateWordTimestamps(text, (audio.length * 8) / 128000);
  }
}

function estimateWordTimestamps(
  text: string,
  totalDuration: number
): WordTimestamp[] {
  const isChinese = /[一-鿿]/.test(text);
  const words = isChinese
    ? text.split("").filter((c) => c.trim())
    : text.split(/\s+/).filter((w) => w);

  const durationPerWord = totalDuration / Math.max(words.length, 1);

  return words.map((word, i) => ({
    word,
    startSeconds: i * durationPerWord,
    endSeconds: (i + 1) * durationPerWord,
  }));
}

registerProvider("openai-tts", () => new OpenAITTSProvider());