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
| Issue | Cause | Solution |
|---|
| Circular dependencies | Wrong layering | Services depend on client, never reverse |
| Cold start latency | Client initialization | Pre-warm in server startup |
| Memory pressure | Unbounded audio cache | Set maxSizeMB on cache |
| Type errors | SDK version mismatch | Pin SDK version in package.json |
| Frequent 429s | Concurrency above plan limit | Lower maxConcurrency in config |
| Missing API key | ELEVENLABS_API_KEY unset | Export it before loadConfig() runs |
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);
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
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.
