TTS vendor survey for a multi-tenant inbound receptionist. The conclusion is counterintuitive: the binding constraint is concurrency, not price or quality, and it disqualifies the two vendors most people would pick first.

The Binding Constraint: Concurrency Caps

ElevenLabs caps at 15 concurrent requests even on the $990/mo Business tier

That is identical to the $299 tier. You cannot buy your way out of it self-serve.

Vendor / tierConcurrent requestsSelf-serve?
ElevenLabs Pro ($99/mo)10yes
ElevenLabs Scale ($299/mo)15yes
ElevenLabs Business ($990/mo)15 — no increaseyes
Cartesia Scale15yes
Rime Starter20yes
Deepgram PAYG45yes, no sales call
Deepgram Growth60yes, no sales call

For inbound telephony, one concurrent request ≈ one caller currently being spoken to. A 15-cap means a hard ceiling somewhere around 15 simultaneous talking agents across all tenants. For a multi-tenant SaaS that is not a growth constraint, it is a launch constraint.

This disqualifies ElevenLabs and Cartesia regardless of their quality

Both are excellent models. Neither can back a self-serve multi-tenant inbound telephony product. Raising the cap requires an enterprise sales motion, which means the cap is a business-model decision, not a temporary limit — plan around it rather than expecting it to move.

Deepgram Aura-2 is the pick, and the deciding factor is a quota page, not an audio sample.

Do Not Pick TTS on Arena Rank

The Artificial Analysis Speech Arena top-4 spread is statistically noise

Blind human preference, 76 models, June 2026: the top 4 span 24 ELO with ±16 confidence intervals. Those intervals overlap. Ranking #1 vs #4 is not a measurable quality difference.

Pick on the things that actually vary and actually matter on a phone line:

  1. Concurrency cap (see above) — hard gate.
  2. μ-law 8kHz support — if the vendor can’t emit it natively you pay a transcode on every frame. See voice-agent-telephony-carriers-2026-07-21.
  3. Latency consistency (IQR), not P50 — see below.

IQR Matters More Than P50 on Phone Calls

Callers notice jitter more than uniform slowness

VendorP50 TTFBIQR
Cartesia188ms100ms
ElevenLabs288ms28ms

Cartesia is 100ms faster at the median and feels worse, because response timing varies unpredictably turn to turn. A consistently-288ms agent reads as “a bit slow”; a 188±100ms agent reads as “glitchy.”

Design implication: benchmark candidate vendors on spread, and treat a tight IQR as worth ~100ms of median.

Skepticism flag — these numbers come via an interested party

The Coval-derived TTS latency figures reach us through Gradium’s blog post, and Gradium tops that benchmark. The raw dashboard at benchmarks.coval.ai/tts is JS-rendered and was not extractable at research time. Re-derive in a browser before treating any specific millisecond figure as load-bearing. [measured, second-hand, conflicted]

Pricing per Minute of Speech

Normalized at ~900 characters per minute of speech.

Vendor / model$/min of speech
Inworld$0.009
Deepgram Aura-20.027
Cartesia$0.034
ElevenLabs Flash v2.5$0.045
ElevenLabs Multilingual v2 / v3$0.090

Note this is per minute of agent speech, not per minute of call. In a 3-minute call the agent typically speaks ~1.1 minutes — see voice-agent-cost-model-2026-07-21 for the conversion.

ElevenLabs v3: The Expressiveness You Want Is Exactly the Thing You Can’t Have

v3 is batch-only, NOT realtime

ElevenLabs’ official docs direct realtime users to Flash v2.5. Turbo v2.5 is deprecated. So the realtime-capable ElevenLabs option is Flash v2.5 at $0.045/min — nearly 2x Aura-2, with a 15-concurrency cap.

The frustrating part: v3’s expressive audio tags[hesitates], [stammers], [pauses], [laughs], [sighs] — are precisely what a human-sounding receptionist would want, and they are precisely unavailable because v3 cannot stream. The most human-sounding output in the ElevenLabs lineup is structurally locked out of live phone calls.

Rime — The Only Spontaneous-Speech Vendor

Rime is the only vendor found that is explicitly engineering spontaneous speech rather than treating naturalness as a byproduct of model scale:

  • Arcana — models filler words, pauses, and prosodic stress.
  • Mist v2 — adds backchanneling affirmations (“mhm”, “right”) and breathing.
  • Trained on conversational rather than scripted speech — this is the structural difference, not a post-hoc effect.

For a receptionist, backchanneling is disproportionately valuable: it’s what makes a caller feel heard while they’re still talking.

BLOCKER — Rime μ-law 8k support is unverified

[unverified] at research time. If Rime cannot emit μ-law 8kHz natively it’s out, regardless of how good the spontaneity modeling is. This is open item #1 in voice-agent and is cheap to close — one API call.

Rime Starter concurrency is 20, which is above ElevenLabs/Cartesia but below Deepgram.

Dead / Avoid

PlayHT / Play.ai is dead

Meta acqui-hired the team. API shut down 2025-12-31. Any tutorial, blog post, or comparison table still listing PlayHT is stale — treat its presence as a staleness signal for the whole document.

Decision

CriterionWinnerMargin
Concurrency (the gate)Deepgram — 45/60 self-serve3–4x the field
PriceInworld ($0.009), then DeepgramDeepgram is mid-pack but adequate
μ-law 8k nativeDeepgramRime unverified
QualityStatistically indistinguishable in top 4Not a differentiator
Spontaneity featuresRimeBlocked on μ-law

Ship on Deepgram Aura-2 with pooled warm WebSocket connections (worth ~250ms — see TTS Connection Warmth Is Worth ~250ms). Revisit Rime once μ-law 8k is confirmed; its spontaneity modeling is the only genuine product differentiator in the TTS field.

Sources

  • Artificial Analysis Speech Arena, June 2026 — 76 models, blind human preference [measured]
  • Coval TTS benchmark, via Gradium blog — TTFB / IQR figures [measured, second-hand from a vendor that tops the benchmark]
  • Vendor pricing and concurrency pages — ElevenLabs, Cartesia, Deepgram, Rime, Inworld [vendor]
  • ElevenLabs official docs — v3 batch-only, Flash v2.5 for realtime, Turbo v2.5 deprecated [vendor]