ElevenLabs · voice
ElevenLabs integration: voice in your product
ElevenLabs integration means your system sends text in and gets audio back, which you store or stream in the product. It pays off when the sound is part of the goods: a spoken guide, a phone message, training, accessibility. It does not pay off as automatic readout of your invoice PDF. Voice licence, cloning a colleague, caching and Danish street names are what break projects — not the HTTP call. A greeting that hits the API 8,000 times is a spend problem you should have seen in the brief. GDPR: the text you send is still personal data if it names a customer.
A phone queue is two integrations: TTS here, and a voice layer such as Twilio. Speech the other way — audio to text — is Whisper. AI features we actually build.
What you must have under control
- Voice licence and “may we clone a colleague’s voice”
- Caching: the same greeting must not hit the API 8,000 times
- Language and pronunciation of Danish street names — test it
- A spend cap per user and per day
Prompt
Menneske godkender før udsendelse
Text
Lock the wording. Minimise names you do not need in the audio.
Synthesise
Call ElevenLabs. Store file, voice id and version.
Cache
Reuse when the text is the same. Count spend.
What to do now
Say whether the sound is the product or decoration. We do not build decoration. Product — guide, message, accessibility — we can put in as one integration. Write what the sound must do.
FAQ about ElevenLabs integration
Can we use it in a phone queue?
Yes, via your voice layer (Twilio and others). That is two integrations: TTS + telephony.
May we clone the director’s voice?
Only with consent and a written agreement. A voice is personal data and an HR topic.
Why cache?
The same sentence should not be paid for 8,000 times. Cache is both price and latency.
Do Danish place names work?
Some. Test street names and company names before you promise readout in production.