Short answer / recommendation:
Use ElevenLabs’ API for generation, keep each video’s narration as small, versioned chunks (chapters/scenes), generate audio via a background worker with concurrency limits, cache on a content+voice+SSML hash, and instrument per-request cost metrics (character count or duration × unit price). Use CI gates that run SSML/format linting and audio smoke tests (durations and silence detection) before release.

Why this works
ElevenLabs gives high-quality voices and SSML control, but you’ll hit rate limits, costs, and occasional re-renders at scale. Treat TTS as a deterministic asset pipeline: small, idempotent jobs you can retry, cache, and audit. This makes builds reproducible and CI-friendly.

Decision criteria (pick priorities first)
- Latency vs cost: real-time preview needs more parallelism and cost; batch overnight cheaper.
- Fidelity vs human performance: ElevenLabs is great for scale, but hire voice actors for premium emotional reads.
- Team skill & infra: if you have dev resources, implement caching, retries, and CI hooks; otherwise use managed orchestration.
- Output granularity: chapter-level chunks reduce re-rendering; sentence-level gives max re-use but increases API calls.

Practical checklist (implement this order)
1) Input normalization: produce a sanitized script with placeholders for names, numbers, and SSML-safe escaping.
2) Chunking strategy: decide chunk size (scene/chapter recommended). Store mapping: video_id → chunk_ids.
3) Voice/version map: maintain a voice registry (voice_id, params, profile_version).
4) Request worker: enqueue TTS jobs (worker pool) with concurrency cap and exponential backoff retries.
5) API request builder: include SSML sections where needed; validate against your SSML whitelist before sending.
6) Caching: key = hash(text + voice_id + SSML + sample_rate + other params). If hit, reuse asset and skip call.
7) Stitching: render chunks to WAV/MP3 and stitch with consistent sample rates; use small crossfades for natural joins.
8) Cost instrumentation: record character_count or seconds_per_clip and per-request cost returned by API; aggregate per-video and per-client.
9) CI gates: SSML lint, duration checks, silence detection, and a small audio-diff (or perceptual hashing) against golden samples for regression.
10) Monitoring & alerts: failures, high-latency, cost spikes, and voice quality complaints.

Best-for / Avoid-if
- Best for: high-volume agencies that need consistent, quickly-updatable voices and programmatic control over prosody and timing.
- Avoid if: you need actor-level emotional nuance on every script or strict non-cloud audio policies (use in-house recordings).

Implementation tips
- Use scene/chapter chunking to minimize re-rendering when edits occur.
- Store metadata (prompt, SSML, voice_id, cost) alongside audio files for auditing and refunds.
- Use ChatGPT to draft/inject SSML or to produce shot-to-voice mapping rules if you want automated script preprocessing.

If you want, I can draft a sample worker + caching pseudocode and a CI checklist tuned to your exact monthly volume, team size, and target latency.