Imported from jeremylongshore/tons-of-skills-marketplace (plugins/saas-packs/elevenlabs-pack/skills/elevenlabs-reference-architecture/SKILL.md). Install upstream with npx skills add jeremylongshore/tons-of-skills-marketplace --skill elevenlabs-reference-architecture. Copyright stays with the author (MIT).

ElevenLabs Reference Architecture

Overview

Production-ready architecture for ElevenLabs TTS/voice applications. Covers project layout, service layers, caching, streaming, and multi-model orchestration. The full code for each layer lives in references/ so this file stays a navigable map; drill into a reference file when you need the exact implementation.

Prerequisites

• Understanding of layered architecture patterns
• ElevenLabs SDK knowledge (see elevenlabs-sdk-patterns)
• TypeScript project with async patterns
• Redis (optional, for distributed caching)
• Auth: an ElevenLabs API key exported as ELEVENLABS_API_KEY (read by the config layer). This is the only ElevenLabs credential — your app's own request auth (middleware/auth.ts) is separate and unrelated.

Instructions

Build the service in six layers. Each step below is the high-level move; the verbatim code and diagrams are in the linked reference files.

Step 1: Lay out the project

Split the codebase into elevenlabs/ (client, config, models, errors, types), services/ (tts, voice, audio, cache), api/ (routes + middleware), queue/, and monitoring/. See the full project tree.

Step 2: Configuration layer

Define an environment-aware ElevenLabsConfig — dev uses the cheap/fast eleven_flash_v2_5 and small output format; production uses eleven_multilingual_v2 at higher quality, more concurrency, and a larger cache. loadConfig() merges the per-environment defaults with ELEVENLABS_API_KEY. Full interface and ENV_CONFIGS: implementation walkthrough.

Step 3: TTS service layer

Wrap the SDK client in a TTSService that owns a singleton client and a p-queue sized to maxConcurrency (this is what prevents 429s). generate() supports both streaming and buffered convert, logs latency, and routes errors through classifyError. generateLongText() splits on sentence boundaries under the 5000-char limit to preserve prosody. Full class: implementation walkthrough.

Step 4: Voice management service

A VoiceService over the client for list/clone/get-settings/update-settings/delete, with category filtering (premade / cloned / generated). Full class: implementation walkthrough.

Step 5: Wire the data flow

Requests flow Client → API layer → Cache/TTS/Voice services → queue → singleton SDK client → ElevenLabs REST/WS endpoints. See the data flow diagram.

Step 6: Health check composition

Compose a /health route that runs connectivity, quota, and cache checks with Promise.allSettled, returning healthy / degraded / unhealthy (degraded once quota exceeds 90%). Full function: implementation walkthrough.

Every architectural choice (singleton client, p-queue, LRU-vs-Redis, sentence splitting, environment-based model selection, HTTP-vs-WS streaming) and its rationale is tabulated in the architecture decisions table.

Output

Applying this skill produces a layered service scaffold, not a single file:

• A directory tree matching the project structure.
• An environment-aware config module resolving dev/staging/production defaults.
• A TTSService (queued, retry-aware, streaming-capable) and a VoiceService.
• A /health route returning { status, services, timestamp } where status is healthy, degraded, or unhealthy.
• At runtime, generate() returns a Buffer (or a ReadableStream when streaming: true); generateLongText() returns Buffer[], one per chunk.

Error Handling

Examples

Generate speech through the service layer:

const tts = new TTSService();
const audio = await tts.generate("Hello from production.", {
  voiceId: "21m00Tcm4TlvDq8ikWAM",
});

Stream a long article with prosody-preserving chunking:

const chunks = await tts.generateLongText(longArticleText);
// chunks: Buffer[] — concatenate or pipe in order

For the complete, runnable layers behind these snippets — config, full TTSService, VoiceService, and the /health composition — see the implementation walkthrough. For the project tree, data flow, and decision rationale, see architecture.md.

Resources

• ElevenLabs API Reference
• ElevenLabs SDK Source
• p-queue
• LRU Cache
• Implementation walkthrough — full service code
• Architecture reference — project tree, data flow, decisions

Next Steps

Start with elevenlabs-install-auth for setup, then apply this architecture. Use elevenlabs-core-workflow-a and elevenlabs-core-workflow-b for feature implementation.