A voice agent for European callers has two hard requirements before any latency number matters: the speech API has to stream over a connection that stays open between turns, and the audio has to stay in the EU. The second one is where providers differ most. Some keep traffic in Europe on every plan, others only inside an enterprise contract with a separate account and key.
This Article compares eight streaming TTS APIs on exactly those points, explains why their published latency figures cannot be compared with each other, and shows what a generated minute costs.
This is the engineering companion to our overview of European AI voice model providers and to Best Text-to-Speech APIs for Developers in 2026.
The author works at KugelAudio, one of the eight providers compared. KugelAudio facts are first-party disclosures; competitor facts come from each provider's own documentation or pricing page. This is not a benchmark: competitor latency figures were not reproduced, enterprise terms are often negotiated, and no winner is declared.
Key takeaways
- Model processing time, server-side time to first audio, client first byte and first playable audio are different measurements and cannot be compared across vendors.
- WebSocket streaming saves the handshake after the first turn and accepts LLM text as it is generated; several providers do the same over HTTP/2 or gRPC style streaming.
- "EU endpoint" can mean EU hosting included on every plan, an enterprise-only isolated environment, or the cloud region of your resource. ElevenLabs and Cartesia offer EU residency to Enterprise customers; KugelAudio runs EU-only by default on every plan for customers in the EU and EEA.
- Converted at 825 characters per minute, the documented list prices here span roughly one to seven cents per minute. Whether EU residency is included at that price, and what committed volume costs, changes the comparison more than the list price does.
Last checked 10 November 2026.
What does "low latency" mean for a streaming TTS API?
Low latency means a short interval between your application having text to speak and the listener hearing the first audio; vendor figures usually cover a smaller slice. Microsoft's Azure Speech documentation names them: "first byte service latency" (request received to first chunk leaving the service), "first byte client latency" (start of synthesis to first chunk at the client, network included) and "finish latency" (until all audio has arrived). Its stated ideal is a user-experienced latency close to one network round trip plus the service's first-chunk latency.
Headline numbers usually leave out connection setup (a TCP handshake, a TLS handshake and an HTTP upgrade, each at least one round trip), the round-trip time itself (much shorter from a Frankfurt server to an endpoint in the same region than across the Atlantic), text normalization and resampling, and decoding: a compressed format may need several frames before it produces sound, and your player may buffer before it starts. Most "model latency" claims describe only inference up to the first audio frame. In a voice agent, all of it follows end of turn, speech recognition and the language model; our performance and capacity planning guide covers instrumenting that chain.
What the published numbers actually measure
Each published figure has its own boundary:
- Mistral documents about 90 ms of model processing for Voxtral TTS, but about 0.8 s to first audio at its API for
pcmand about 3 s formp3: the same model, two boundaries, and a gap caused by the output format alone. - ElevenLabs lists Flash v2.5 at "~100ms", a median "excluding application & network latency".
- Cartesia states "first audio in under 90ms" for Sonic, without describing client location, network path or method.
- Deepgram publishes "time to first audio, as low as" 80 ms for its newer Flux TTS model at 24 kHz, and no figure for Aura-2.
- KugelAudio publishes a 91 ms median time to first audio for Kugel 3 over WebSocket from an EU data centre, network round trip included, connection setup excluded (first-party, not independent). Kugel 3.5, the current default model, is built on Kugel 3; its own latency is not published.
- Azure Speech, Google Cloud Text-to-Speech and Amazon Polly publish no number in the documentation cited here.
A figure that excludes the network cannot be compared with one that includes it, and a median from one data centre says nothing about your p99 from another.
Why does WebSocket streaming matter for voice agents?
A persistent WebSocket pays the handshake once, so later replies do not wait for TCP, TLS and an HTTP upgrade, and it accepts text while the language model writes it. Cartesia's endpoint comparison agrees: with a WebSocket "you pay connect cost once", while on HTTP the connection overhead "is often on the same order as TTFB for the audio itself".
"Streaming" covers four shapes. Output streaming over HTTP takes complete text and returns audio in chunks or server-sent events. A single-request WebSocket sends whole texts over a reused socket. Input streaming over WebSocket accepts text fragments and lets the server decide when to synthesize, which is what an LLM-driven agent needs. Google Cloud and Amazon Polly stream bidirectionally without WebSocket. The table below maps providers to shapes.
Whatever the protocol, test whether partial text is accepted, whether the connection is reused across turns, and whether a turn can be cancelled without closing it.
How do chunking and flushing change time to first audio?
How you hand text to the server often changes time to first audio more than the choice of model, as KugelAudio's documented semantics show. The server synthesizes each sentence when its closing ., ? or ! arrives, and text without a sentence end after 2 seconds without new input. A flush ends the turn, so flushing after every sentence splits one reply into several turns, each paying a fresh model prefill. The documentation recommends sending tokens as they arrive and flushing once per reply. Other providers' flush controls work differently.
The codec decides when the first frame can play. Raw PCM usually plays on arrival, a compressed format may need several frames, and Mistral's pcm and mp3 figures show the scale. A non-native sample rate adds resampling; KugelAudio's native rate is 24 kHz. For telephony, check whether the provider outputs the G.711 codec your carrier expects.
What happens when the caller interrupts?
The stream has to stop quickly, which takes two actions: cancel generation on the server and stop playback locally. On KugelAudio, {"cancel": true} on the socket (or cancel_current() in the Python SDK) abandons the current turn, drops buffered text and is acknowledged with interrupted instead of final, while the WebSocket stays open for the next turn. A few frames already in transit can arrive before the acknowledgement, so stop local playback at once.
For every provider, cancel at the start, middle and end of a long reply and record time to silence, late audio from the cancelled turn, and whether the connection survives. Ask whether cancelled audio is billed.
Which TTS providers offer EU-hosted streaming endpoints?
All eight providers here document some way to keep traffic in Europe, but the access model decides how fast you can start. KugelAudio includes its EU Sovereign Cloud on every plan for customers in the EU and EEA, with no enterprise contract. Deepgram, Azure, Amazon Polly and Google Cloud offer EU endpoints or regions that you select yourself. ElevenLabs and Cartesia offer EU deployments only to Enterprise customers, with a separate account and key. Mistral hosts API data in the EU by default, with documented exceptions.
An EU endpoint is a routing and hosting control, not a compliance outcome: retention, subprocessors, support access and the contract still need review. Voices differ by region: Google restricts some regional endpoints to Neural2 voices, and Polly's London and Zurich regions support a subset of generative voices. Enterprise-only residency usually means a separate account, key and contract before you can test.
How much does a streaming TTS minute cost?
At list price a generated minute costs roughly one to seven cents, but list prices only compare fairly when the same hosting is included. At ElevenLabs and Cartesia the self-serve price does not include EU residency; it is part of an enterprise agreement. The table therefore shows, next to each list price, whether EU residency is part of that price.
Most providers bill input characters and KugelAudio bills generated audio duration, so the last column assumes 825 characters per minute, the Artificial Analysis convention that KugelAudio's pricing configuration also uses; ElevenLabs' pricing page equates 1,000 characters to roughly one minute, and Cartesia states about 750 to 800 credits per minute. That column is arithmetic, not a quoted price.
KugelAudio prices are in euros excluding VAT, the others in US dollars, all without free tiers, plan fees or discounts. KugelAudio's Enterprise plan adds lower rates with committed usage, SLAs and on-premise hosting; the other providers' committed-use terms are negotiated individually and not compared here. Duration and character billing also react differently to speaking rate, pauses and markup. For a fair number, replay a representative month of your own text through each candidate and calculate the all-in cost per successful generated hour, including failed, cancelled and regenerated requests; our developer API comparison gives the formula. Also check whether the plan permits commercial use of the output, whether voice cloning is included, and which concurrency limits apply before you hit 429 responses.
How do you measure streaming TTS latency yourself?
Shortlist on hosting and streaming shape first, then measure from the region where your application will run, not from a laptop, with model, voice, text, codec and sample rate fixed. Time the handshake separately, since users only pay it if your application connects per request. Record per turn:
t0 connection opened (measured separately, once per session)
t1 first text sent
t2 first audio byte received
t3 first decodable audio frame available
t4 playback started
t5 final audio received, or cancel acknowledged
KugelAudio's Python SDK documentation shows the pattern: time KugelAudio.create(..., region="eu") on its own, then stop the first-audio clock at the first audio chunk from client.tts.stream_async.
Repeat each scenario often enough to support the percentile you report, and state the count. Keep cold and warm connections, short and long texts and barge-in separate, and test every shortlisted provider on the same day from the same machine. For concurrent-stream load tests, follow the method in our capacity planning guide.
FAQ
Which TTS API has the lowest latency in Europe?
Public numbers cannot settle it, because each provider measures a different boundary. Measure first playable audio from your own EU region with reused connections and compare p50, p90 and p99.
Do I need a WebSocket for low-latency TTS?
Not strictly. A WebSocket avoids repeated handshakes and accepts partial text, which matters for LLM-driven agents, but Google Cloud and Amazon Polly stream bidirectionally without it, and complete text over a reused HTTP connection can be enough.
Does an EU endpoint make a TTS API GDPR compliant?
No. It controls where requests are routed and processed, not retention, subprocessors, support access or your own obligations. Review the provider's data processing terms and confirm in testing that traffic uses the EU host.
What is the cheapest WebSocket TTS API for commercial use?
It depends on whether EU residency has to be part of the price. Among self-serve list prices that include EU hosting, Azure neural voices in an EU region are at the low end; ElevenLabs and Cartesia need an enterprise agreement for EU residency, while KugelAudio includes it on every plan and lowers the rate with committed usage. Replay a representative month of your own traffic before deciding.
How much latency does a US-hosted endpoint add for European users?
It depends on where your servers run and how traffic is routed, so this Article gives no figure. Measure from your production region against each provider's default endpoint and its EU host.
Does KugelAudio accept text streamed token by token from an LLM?
Yes. The /ws/tts/stream WebSocket buffers text as it arrives and synthesizes at sentence boundaries. Flush once per reply, and on barge-in cancel the turn while keeping the connection open.
Sources
All sources accessed 10 November 2026.
- KugelAudio, Text to Speech product page (languages, Kugel 3 latency and boundary): https://www.kugelaudio.com/en/text-to-speech
- KugelAudio, pricing: https://www.kugelaudio.com/en/pricing
- KugelAudio documentation, Latency: https://docs.kugelaudio.com/latency
- KugelAudio documentation, Regions: https://docs.kugelaudio.com/guides/regions
- KugelAudio documentation, Chunking and per-segment latency: https://docs.kugelaudio.com/streaming/chunking-and-latency
- KugelAudio documentation, Barge-in: https://docs.kugelaudio.com/streaming/barge-in
- KugelAudio documentation index: https://docs.kugelaudio.com/llms.txt
- ElevenLabs, Models: https://elevenlabs.io/docs/overview/models
- ElevenLabs, WebSocket stream-input reference: https://elevenlabs.io/docs/api-reference/text-to-speech/v-1-text-to-speech-voice-id-stream-input
- ElevenLabs, Data residency: https://elevenlabs.io/docs/overview/administration/data-residency
- ElevenLabs, API pricing: https://elevenlabs.io/pricing/api
- Cartesia, Sonic: https://cartesia.ai/sonic
- Cartesia, Compare TTS endpoints: https://docs.cartesia.ai/use-the-api/compare-tts-endpoints
- Cartesia, Regional endpoints: https://docs.cartesia.ai/enterprise/regional-endpoints
- Cartesia, Pricing: https://www.cartesia.ai/pricing
- Deepgram, Regional endpoints: https://developers.deepgram.com/reference/regional-endpoints
- Deepgram, Pricing: https://deepgram.com/pricing
- Deepgram, Text to speech product page: https://deepgram.com/product/text-to-speech
- Microsoft, How to lower speech synthesis latency: https://learn.microsoft.com/en-us/azure/ai-services/speech-service/how-to-lower-speech-synthesis-latency
- Microsoft, Supported regions for Azure Speech: https://learn.microsoft.com/en-us/azure/ai-services/speech-service/regions
- Microsoft, Azure retail prices API (Azure Speech, West Europe): https://prices.azure.com/api/retail/prices
- Google Cloud, Bidirectional streaming synthesis: https://docs.cloud.google.com/text-to-speech/docs/create-audio-text-streaming
- Google Cloud, Regional endpoints: https://docs.cloud.google.com/text-to-speech/docs/endpoints
- Google Cloud, Text-to-Speech pricing: https://cloud.google.com/text-to-speech/pricing
- AWS, Polly bidirectional streaming: https://docs.aws.amazon.com/polly/latest/dg/bidirectional-streaming.html
- AWS, Polly generative voices and regions: https://docs.aws.amazon.com/polly/latest/dg/generative-voices.html
- AWS, Polly pricing: https://aws.amazon.com/polly/pricing/
- Mistral, Text to speech documentation: https://docs.mistral.ai/studio-api/audio/text_to_speech
- Mistral, Speech endpoint reference: https://docs.mistral.ai/api/endpoint/audio/speech
- Mistral, Voxtral TTS announcement and price: https://mistral.ai/news/voxtral-tts
- Mistral, Where data is stored: https://help.mistral.ai/en/articles/347629-where-do-you-store-my-data-or-my-organization-s-data
Build with KugelAudio
Put European voice infrastructure into production.
Use the EU endpoint or discuss a customer-operated Kubernetes deployment.