Numbers, read like a person.
It reads an IBAN in the groups you wrote, the way you would on the phone.
90.8%of IBANs and codes read in their groups, like a person would1
Your IBAN is GB29 NWBK 6016 1331 9268 19, due on 15 March, reference 7703 2481.
Text to SpeechKugel 3.5
Fewer mistakes.
More conversations.
Our newest voice model reads numbers, names and one-word replies the way a person would.1
“Yes.” Sounds like yes.
One-word answers come out clear and whole, never clipped.
3%of short replies said wrong, against 6.1% for the next best1
- Yes.
- No.
- Sure.
- Thanks.
One voice. One accent.
A German voice says English words without slipping into another accent.
0%of words spoken in the wrong accent1
Das Meeting zum Quarterly Review ist auf Freitag verschoben, das Update kommt per E-Mail.
Same texts. Same judges. Fewer mistakes.
9 tests against the voices you are comparing.1
| Test | Kugel 3.5 | ElevenLabs v4 Turbo | Cartesia Sonic 3.6 | Deepgram Aura-2 |
|---|---|---|---|---|
| Words pronounced with the wrong accent given the reference voicelower is better | 0% (best) | 4.3% | 0.5% | 12.8% |
| Short words said wronglower is better | 3% (best) | 7.7% | 6.1% | 20.6% |
| Abrupt endingslower is better | 0.7% | 0% (best) | 3.2% | 42.4% |
| Delay before the first wordlower is better | 40 ms (best) | 60 ms | 160 ms | 340 ms |
| 200 ms pauses within 50 mshigher is better | 46.3% (best) | 1.8% | 36% | 15.5% |
| German IPA pronunciationhigher is better | 65.2% (best) | 57.3% | 0.8% | 29.4% |
| English IPA pronunciationhigher is better | 75% (best) | 65.6% | 6.2% | 32.5% |
| Numbers read exactlyhigher is better | 97.1% (best) | 95.4% | 94% | 36.1% |
| Numbers and codes read in their groupshigher is better | 90.8% (best) | 71.9% | 82.6% | 0.9% |
Listeners pick it.
First of 6 in a blind listening test.2
ModelWin rate
- 1KugelAudioScore 2678.0%
- 2ElevenLabs Multilingual v2Score 2562.2%
- 3ElevenLabs v3Score 2165.3%
- 4CartesiaScore 2159.1%
- 5VibeVoiceScore 1028.8%
- 6CosyVoice v3Score 914.2%
Listeners heard a reference voice, then picked the model that sounded more human and closer to the original. German samples: neutral speech, shouting, singing and slurred speech.
Data Sovereignty
Your data stays in Europe
Built, trained and hosted in the EU.
EU Servers
Hosted in Europe
GDPR
Compliant
No training
Never on your content
EU/EEA customers
Sovereign Cloud
Hosting by EU companies only. AWS, Google Cloud and Microsoft Azure are not engaged under the Sovereign Cloud agreement at all. No provider with a US parent processes your audio, text or voice references, so no US CLOUD Act order can reach them.
Available to EU/EEA customers
Enterprise
On-Premise
Maximum security on your own hardware. The models run in your data centre, behind your firewall, so audio, text and voice references never leave your infrastructure.
Same SDK and API as the hosted service, so integrations do not change.
Plugs right in.
LiveKit, Pipecat, Vapi and Cognigy. Or point your ElevenLabs code at us.
Make your first request- LiveKit AgentsOfficial Python and TypeScript plugin: pass it to AgentSession, prewarm, one WebSocket per agent.
- PipecatDrop-in TTS service for Python pipelines; time to first audio shows in Pipecat’s metrics.
- VapiBuilt-in custom TTS endpoint, no proxy server.
- CognigyRegister once as a custom speech vendor in Voice Gateway.
- Coming from ElevenLabsCompatible API: change the base URL, voice ID and output format.
Tech specs
kugel-3.5- Models
- kugel-3.5, the default; kugel-3 when you set it
- Languages
- 39, one multilingual model
- Time to first audio, Kugel 3
- 91 ms median; WebSocket streaming, network round trip included, connection setup excluded3
- Streaming
- Text in, audio out; REST or WebSocket
- Voice cloning
- From your reference audio
- Audio output
- 24 kHz native; 8, 16, 22.05, 24, and 44.1 kHz; 8 kHz for telephony, no resampling step
- Input
- Up to 10,000 characters per request
- Hosting
- Built, trained and hosted in the EU
- Older model IDs
- kugel-2.5, kugel-2-turbo, kugel-2, kugel-1, and kugel-1-turbo run on kugel-3.5; kugel-1 and kugel-1-turbo keep a 5,000 character limit
- Docs
Streaming in both directions
Flush once per turn; the server groups text at sentence boundaries.
flush - Docs
Voice cloning
Clean recordings of a single speaker work best.
- Docs
39 languages, one model
Name the language when it differs from the voice’s primary language.
language="de" - Docs
Built-in text normalization
Numbers, dates and currencies become words before synthesis.
- Docs
Custom pronunciation with IPA
IPA between slashes, or a pronunciation dictionary for every request.
/ˈkuːɡl̩/ - Docs
Spelling for codes and addresses
Read character by character, codes in groups of 4.
<spell>GB29NWBK</spell> - Docs
Pauses where you want them
A break tag inserts silence at that exact point.
<break time="400ms"/> - Docs
Word-level timestamps
Timing for every word, sent after its audio and never holding it back.
Three habits keep first audio fast
Latency depends on region, network and load. Measure your own deployment.
Connect at startup
No caller waits for the handshake.
Name the language
The server does not detect it from the text.
Flush once per turn
Every flush forces a fresh model prefill.
Questions about Kugel 3.5
Which model do I get?
Kugel 3.5, the default when you send no model ID. kugel-3 stays available: set model_id to kugel-3.
How does Kugel 3.5 compare with ElevenLabs, Cartesia and Deepgram?
On the same texts and judges, it leads on 8 of 9 KugelBench tests, including numbers, codes and accents.
How natural does it sound?
Listeners ranked KugelAudio first of 6 models in a blind test on German speech, with a 78.0% win rate. The test predates Kugel 3.5.
How fast is it?
Kugel 3.5 is built on Kugel 3, the fastest model in our latency benchmark; its own latency is not published yet. Region, network and load change latency, so measure your own deployment.
Can I use my own voice?
Yes. Voice cloning creates a voice from clean reference audio of a single speaker, and you use it like a ready voice.
Where is my data processed?
In the EU: KugelAudio is built, trained and hosted there, is GDPR compliant and never trains on your content. EU Sovereign Cloud is included for EU/EEA customers, and Enterprise can run on premise.
I use kugel-2.5 or an older model ID. What changes?
It now runs on kugel-3.5, and usage and billing show the ID you sent. Send kugel-3 to stay on Kugel 3.
Hear it read your script
Create an account and paste the sentence your agent says most.
- 1Kugel 3.5 against ElevenLabs v4 Turbo, Cartesia Sonic 3.6, and Deepgram Aura-2 on 9 tests. KugelBench, Kugel 3.5 on 08.10.2026, the other models between 29.09.2026 and 06.10.2026. Every model spoke the same texts and was scored by the same automatic judges. Kugel 3.5 leads 8 of them; ElevenLabs leads abrupt endings.
- 2Blind listening test of 6 models on German speech, OpenSkill ranking; run before Kugel 3.5 and published as KugelAudio.
- 3Kugel 3: 91 ms median time to first audio over WebSocket streaming from an EU data centre, network round trip included, connection setup excluded. Kugel 3.5 is built on Kugel 3; its own latency is not published here.