ElevenLabs makes the voices listeners rank highest, and ElevenAgents is a hosted agent platform around them. It suits teams that want both from one vendor. Its compliance terms sit on Enterprise. Dograh, an open-source AI voice agent platform you can self-host or run on Dograh Cloud, can use those voices with your own key.
Key Takeaways
- Eleven v4 models top the Speech Arena, while Flash v2.5 sits outside the top 50.
- Compliance terms such as a HIPAA agreement need an Enterprise plan.
- Teams can keep ElevenLabs voices on an open-source platform they control.
ElevenLabs is known for its voices, but a phone agent also needs a platform, a fair price and compliance terms. We cover which voice model you actually run, ElevenAgents on a call, the Enterprise terms, cost, who it suits and keeping its voices inside Dograh. It belongs with our guide to ElevenLabs alternatives.
The voices: which model you actually run
ElevenLabs' voices are its strongest card, and the model you pick decides how much of that you get. On the Artificial Analysis Speech Arena, where listeners vote blind between two clips, Eleven v4 and Eleven v4 Turbo hold the top two places. Eleven v3 sits at 23rd.
Flash v2.5 is a different story. It is the fast model many phone agents run by default, and it sits outside the top 50 on the same board. Speed matters on a call, but so does sound, so test both on your own script before you decide. A review that says "ElevenLabs sounds the best" is only true for the model you actually run.
Transcription is strong too. Scribe v2, ElevenLabs' batch model, ranks fourth on the Artificial Analysis word error rate index. An agent runs Scribe v2 Realtime, the streaming version, and that batch result says little about it, so test the streaming model on your own calls.
Voice cloning and a large voice library round it out. You can clone a voice from a recording and use it in an agent, which matters if your brand already has a recognisable voice.
ElevenAgents on a phone call
ElevenAgents is a full hosted agent platform that arrives already assembled. The ElevenAgents overview lists ElevenLabs' own fine-tuned speech-to-text, a proprietary turn-taking model, a choice of supported language models or your own, and the voices. Around them sit a visual workflow builder, tools, a knowledge base, built-in testing and evals, native Twilio, SIP trunking and SDKs for iOS, Android, React Native, Unity and React.
The phone features are deeper than most hosted platforms offer. Transfers can be conference, blind or SIP REFER, and callers can answer with the keypad while the agent can press keys itself. Plans run from 4 concurrent calls on Free to 40 on Business, and burst pricing lets a plan take up to three times its usual concurrency.
On latency, read the numbers carefully. ElevenLabs quotes how fast a voice model returns its first audio, and that is not what the caller hears. A caller's wait also includes deciding the caller has finished, transcription, the language model, any tool call and the phone network. Our post on barge-in, endpointing and turn-taking explains where that time goes. Measure the gap from the caller stopping to the agent speaking on a real phone line.
Compliance and hosting sit on Enterprise
The compliance terms most regulated buyers need are on the Enterprise plan. ElevenLabs' HIPAA terms make the business associate agreement (BAA) available only to Enterprise customers, and patient data also needs Zero Retention Mode turned on.
That mode narrows the stack. Only allowlisted language models can run with it, such as Gemini, Claude and OpenAI GPT under a BAA, ElevenLabs-hosted Qwen3.5 and GLM-5.2, or a custom OpenAI-compatible model after telling ElevenLabs Support. Among the realtime speech-to-speech models, only Eleven Realtime v1 Max works there.
Hosting follows the same line. Agents run on ElevenLabs' cloud by default, and EU, India or Singapore data residency is an Enterprise feature. ElevenLabs' private deployment runs ElevenAgents inside your own AWS or GCP account, arranged through sales, with the v2 and v2.5 voices and Scribe V2. The newest Eleven v3 and v4 voices are not part of it.
Open Source Alternative to Vapi / Retell
Self-hosted voice agent platform — no per-minute fees
dograh-hq/dograh
Star on GitHub
Cost and where it falls short
ElevenAgents bills by the call minute. Plans run from a free plan with 15 minutes to Business at $990 a month for 12,375 minutes, extra minutes cost $0.08, and calls above your concurrency limit cost $0.16 a minute. The language model bills on top, and your carrier bills its own minutes. Our breakdown of ElevenLabs pricing works through a full month.
The weak points come from how the product is built. The platform code is closed, and the public GitHub holds SDKs and starter kits. Inside an agent, the speech-to-text and the voice are ElevenLabs' own, so you cannot pair its turn-taking with another vendor's transcription. Your own language model plugs in through an OpenAI-compatible endpoint that ElevenLabs' cloud calls, and it cannot be paired with the Realtime models.
The bill is also split three ways for the same call: ElevenLabs' minute, the language model and the carrier. None of that is hidden, but it makes a monthly forecast harder than one per-minute rate suggests.
Who ElevenLabs suits, and who should look elsewhere
ElevenLabs suits a team that wants the top-ranked voices and a ready agent from one vendor, with no servers to run. It fits brands where the voice is part of the product, teams that need keypad input or SIP REFER transfers, and buyers already on an Enterprise contract for the BAA and residency.
Look elsewhere if you need the platform's code, a free choice of speech-to-text and language model on every call, or patient data on your own servers without an Enterprise contract. The same goes for teams that want the newest voices inside a private deployment. If open voices on your own hardware are the goal, our open-source ElevenLabs alternative post compares them.
Join the Dograh Community
Dograh is an OSS alternative to Vapi. Join our Slack community for queries, releases, best practices & community interactions.
Keeping ElevenLabs voices inside Dograh
You can keep ElevenLabs' voices and own the platform around them. The platform handles the call flow, tools and telephony, and the voice becomes one pick in a speech and language layer you assemble yourself. ElevenLabs is a built-in voice provider in Dograh, from Flash v2.5, the default, up to Eleven v4 and v4 Turbo. You paste the voice ID from your Voice Library, cloned voices included, add your own ElevenLabs key and, in the voice settings, can point the base URL at an ElevenLabs Data Residency endpoint.
| ElevenAgents | Dograh with ElevenLabs voices | |
|---|---|---|
| Code | Closed platform, public SDKs | Open source, BSD-2-Clause |
| Voices | ElevenLabs only | ElevenLabs on your key, or other providers and local voices |
| Speech-to-text | ElevenLabs' own | Scribe v2 Realtime or another provider |
| Language model | Supported models or a custom endpoint | Any supported provider or a local model |
| Hosting | ElevenLabs' cloud, private deployment on Enterprise | Self-host or Dograh Cloud |
| BAA | Enterprise only | None with Dograh when self-hosted, ElevenLabs needs one if its voices handle patient data |
| Keypad | Reads and plays touch tones | Callers answer by voice |
| Transfers | Conference, blind, SIP REFER | Transfer Call on Twilio, Telnyx, Plivo, Cloudonix and Asterisk ARI |
With your own key, ElevenLabs still processes the speech it voices or transcribes, so plan your data rules around that. The voice bill stays on your key either way. Self-host Dograh and the platform adds no per-minute fee, leaving servers, model keys and phone minutes, or pay a 1¢-a-minute platform fee on Dograh Cloud.
Agents are built in a visual workflow builder, and when a caller asks for a person, Dograh rings that colleague first and puts the caller through once they pick up. Our Dograh vs ElevenLabs comparison goes through the two platforms row by row.
Glossary
- Speech Arena
- Artificial Analysis's leaderboard where listeners vote blind between two voice clips. Ranks move as votes come in, so name the exact model whenever you quote one.
- Zero Retention Mode
- An ElevenAgents setting that stops ElevenLabs from keeping call content. Patient data needs it on top of an Enterprise BAA, and it limits which language models an agent can use.
- Time to first audio
- How fast a voice model returns the first sound after it gets text. It is one slice of a caller's wait, which also includes turn detection, transcription and the language model.
- Burst pricing
- ElevenAgents' rate for calls above a plan's concurrency limit, up to three times the usual number of calls, charged at $0.16 a minute instead of $0.08.
