ElevenLabs vs Cartesia vs PlayHT for UK Voice Agents

We A/B-tested two identical outbound scripts — same call flow, same LLM, same campaign list — using ElevenLabs Turbo v3 on one number pool and Cartesia Sonic-2 on the other. Answer rates were identical. But across a 300-call sample, the ElevenLabs pool added 94ms to average turn latency. At 200 calls per day, that 94ms compounds: callers end conversations 11% faster before the qualification question lands. The TTS provider is not an afterthought in a latency-constrained stack. And there is the other side: when we played a convincing Geordie or a Glaswegian voice to a UK prospect rather than a neutral mid-Atlantic accent, callback rates shifted. Voice quality is real. So is the price difference — ElevenLabs costs four times what Cartesia costs at 10,000 minutes per month.

ElevenLabs, Cartesia, and PlayHT: what each TTS provider actually does at the API level

ElevenLabs operates on an HTTP REST API with an optional WebSocket endpoint. Turbo v3 is their latency-optimised model — it trades some expressive range for speed and sits at roughly 134ms time-to-first-byte from London-proxied requests. The ElevenLabs streaming API supports chunk-by-chunk text input so you can begin audio generation before the LLM has finished its full response.

Cartesia built WebSocket-first from the start. Their Sonic-2 model keeps the connection open across turns, streams audio within 68ms, and lets you send the next text chunk while the previous audio is still playing. There is no handshake cost on turns two through ten of a conversation — that is why Cartesia consistently wins on raw TTFB.

PlayHT sits between the two. Their PlayDialog model supports both REST and WebSocket streaming, with TTFB benchmarks around 108ms from UK infrastructure. PlayHT's differentiator is expressive prosody on longer utterances, but that expressiveness costs latency on the short back-and-forth turns that define real-time voice agents.

Here is the configuration block we use in our voice AI architecture to switch providers at runtime without touching call-flow logic:

{
  "tts_provider": "cartesia",
  "cartesia": {
    "model_id": "sonic-2",
    "voice_id": "a167e0f3-df7e-4d52-a786-6f5c8ee9f399",
    "output_format": "pcm_16000",
    "stream": true,
    "language": "en"
  },
  "elevenlabs": {
    "model_id": "eleven_turbo_v3",
    "voice_id": "ErXwobaYiN019PkySvjV",
    "output_format": "pcm_16000",
    "stream": true,
    "optimize_streaming_latency": 4
  },
  "playht": {
    "model": "PlayDialog",
    "voice": "s3://voice-cloning-zero-shot/775ae416-49bb/manifest.json",
    "output_format": "raw",
    "sample_rate": 16000,
    "stream": true
  }
}

This pattern powers our voice AI document analysis build, where we hot-swap providers when one endpoint degrades under load.

Voice quality on UK accents: Southern English, Scottish, Welsh, and Northern Irish test results

We ran 48 test utterances — 12 per regional accent — through all three providers using their closest-matching catalogue voices, covering typical outbound openers, qualification questions, and objection-handling turns. Each clip was rated blind by four native speakers per region.

ElevenLabs has a substantially deeper UK catalogue — 40+ regional options versus a handful for each competitor. Cartesia Sonic-2 handles RP and near-RP acceptably but generates audible artefacts on rhotic Scottish vowels. PlayHT PlayDialog delivered the most natural mid-sentence prosody in English RP but collapsed outside Southern English accents.

If you run outbound into Scotland or Wales and want accent-matched voices, ElevenLabs is the only realistic option. If you are running RP on an inbound channel, all three are viable and latency plus pricing should drive the decision.

Latency benchmarks from UK datacentres: time-to-first-byte and streaming onset per provider

We measured TTFB from a London VPS across 200 requests per provider on a Tuesday morning between 09:00 and 11:00 GMT. Inputs were 25-word agent turns — the typical length of a qualification opener.

The 66ms gap between Cartesia and ElevenLabs is what we saw in controlled testing; our production 94ms gap included voice broker overhead and occasional connection re-establishment on ElevenLabs's end.

At 300ms total turn latency — the threshold where callers begin perceiving a lag — those 66ms matter. They leave 234ms for ASR plus LLM first-token versus 168ms. Our post on prompt engineering for sub-second voice agent turns covers what you can recover on the LLM side, but 66ms saved in TTS costs nothing once you have switched provider. If ASR is also contributing latency, our Deepgram vs Whisper vs AssemblyAI for UK voice agents covers the STT side of the same stack.

Pricing at 10k, 50k, and 200k minutes per month: the real annual bill with no credits left out

Pricing as at July 2026. Converted to GBP at £0.80/USD. ElevenLabs bills per character; Cartesia and PlayHT bill per minute. We normalised to minutes using 800 characters per minute — a typical conversational TTS output rate.

These are list prices with no promotional credits. ElevenLabs offers enterprise contracts above 200,000 characters per month. Cartesia has volume tiers above 500,000 minutes. Neither guarantees UK data residency at any tier.

The counter-argument to choosing Cartesia on cost alone is worth stating. PlayHT's own benchmarks show PlayDialog outperforming Sonic-2 on naturalness scores for complex sentences. If your agent script includes multi-step policy reads, that quality difference may justify the premium. A 5% improvement in contact-to-qualification rate across a £4M pipeline outweighs a £100k annual TTS spend — run your own A/B test before committing to volume pricing.

Voice cloning for outbound agents: what works, what sounds robotic, and what UK law covers

All three providers support instant voice cloning from short audio samples: ElevenLabs requires roughly one minute of clean audio, Cartesia and PlayHT work with 30 seconds. Professional cloning — fine-tuned on 30+ minutes of source material — produces materially better results but takes one to three business days per provider.

In testing: Cartesia's instant clone holds up for the first 8–10 seconds of each turn, then shows prosodic drift on longer utterances. ElevenLabs's instant clone maintains consistency longer but the clone model adds 40–60ms of processing time versus a standard voice. PlayHT's instant clone is the most consistent on short outbound turns but shows vowel compression on Scottish source voices.

UK law here is clear. The ICO's guidance on biometric data classes a cloned voice as biometric data under UK GDPR Article 9, requiring explicit informed consent from the voice donor before any processing. You cannot use a clone to misrepresent the identity of a speaker to call recipients — a clone of your named sales director presented as that individual on outbound calls is both a PECR and a UK GDPR problem.

The compliant route: clone a voice actor under a written consent contract, disclose the call as automated at the outset, and store the consent record against the voice asset ID. Two UK financial services clients have run this through legal review without issue.

Streaming TTS and barge-in compatibility: which providers handle interruptions without clipping

Barge-in — where the caller speaks while the agent is talking and the agent must stop mid-sentence — requires the voice broker to cancel an in-flight TTS stream. Each provider handles this differently, and the differences matter more than the TTFB benchmarks in conversational deployments.

Cartesia's WebSocket API includes a cancel frame that halts audio delivery immediately. In testing, audio cut off within 80–100ms of sending the cancel signal — clean enough that callers do not notice a trailing fragment. ElevenLabs requires closing the WebSocket connection and reopening it, which adds 150–200ms of silence before the agent begins listening again. PlayHT's streaming endpoint has no cancel mechanism at all: you must drop the connection, and the audio buffer continues playing for up to 400ms after the caller has started speaking.

That 400ms clip is not marginal. Callers who interrupt and hear the agent continue for nearly half a second before stopping report a mechanical, unresponsive feel — and they hang up sooner. For any outbound campaign where callers commonly talk over the opening line, PlayHT fails this test.

Our post on voice agent barge-in handling covers the broker-side implementation in detail, including how to flush the VAD input buffer simultaneously with the TTS cancel to prevent clipped audio from echoing back into your ASR model.

Data residency and UK GDPR: where each provider processes and stores your audio

None of the three providers currently offer UK data residency. Audio sent to ElevenLabs, Cartesia, or PlayHT leaves UK jurisdiction.

Under UK GDPR, transferring audio to the US requires a valid transfer mechanism. The correct post-Brexit instrument is the UK International Data Transfer Agreement (IDTA), not EU Standard Contractual Clauses. All three providers offer DPAs, but request the UK IDTA addendum explicitly — the default template references EU SCCs. See the ICO's IDTA guidance for the approved wording.

Cartesia's no-retention policy is the cleanest from a data minimisation standpoint — audio processed but not stored reduces exposure in the event of a provider breach.

What changed in 2025–2026: Cartesia Sonic-2, ElevenLabs Turbo v3, and PlayHT PlayDialog benchmarks

Three significant releases between November 2025 and March 2026 changed the calculus for UK voice agent deployments.

Cartesia released Sonic-2 in November 2025. Compared with Sonic-1, it reduced p50 TTFB by 35%, added 16 new languages, and improved Scottish and Welsh accent handling enough to be usable as a generic UK agent voice. The Cartesia Sonic-2 changelog cites a 28% improvement on their internal naturalness benchmark.

ElevenLabs launched Turbo v3 in February 2026, directly targeting Cartesia's latency position. The gain over Turbo v2 is real — roughly 30ms off p50 TTFB and significantly better reliability under sustained load. Turbo v2 dropped frames above 50 concurrent streams; Turbo v3 holds up to 150 in our load tests.

PlayHT deprecated Play3.0-mini in January 2026 and replaced it with PlayDialog. PlayDialog narrows the expressiveness gap with ElevenLabs at longer utterances and adds explicit per-minute pricing, making it easier to forecast costs. At p50, latency is still 40ms behind Cartesia Sonic-2 — PlayHT is not competing on speed.

Our recommendation by use case: high-volume outbound vs inbound triage vs appointment booking

High-volume outbound (100+ calls per day): Use Cartesia Sonic-2. The latency advantage compounds at scale, the price is four times lower than ElevenLabs, and the barge-in cancel implementation is the cleanest of the three. Accept a neutral English RP voice rather than strong regional variation. Pair with TTS caching for repeated phrases — greetings, compliance disclosures, hold messages — to reduce billable minutes further. Our post on TTS caching for voice agents covers the implementation pattern with cache hit rates above 40% for typical outbound scripts.

Inbound triage (quality over volume): ElevenLabs Turbo v3 while you identify which voice performs best with your inbound caller base. Inbound callers are less latency-sensitive — they initiated contact. Once you exceed 10,000 minutes per month with a stable voice preference, evaluate cloning it into Cartesia to reclaim the cost differential.

Appointment booking agents (mid-volume, multi-sentence turns): PlayHT PlayDialog if your booking flow involves long agent utterances — address confirmation, policy readings, multi-step consent scripts. The prosody over multi-sentence turns is the strongest of the three. Hard block: do not use PlayHT if your callers regularly interrupt. The 400ms clip artefact will cost you more in conversation quality than the prosody improvement gains.

Good / Bad / Ugly:

The choice is not which provider is best — it is which provider fits the constraints of your specific deployment. Know your call volume, your accent requirements, and your barge-in rate before committing to a contract.