Measurements

How we measure our voices.

We publish numbers we took ourselves, with the date and the method. Results are mixed: no system leads every column. A human blinded listen is still pending.

Target≈300 msend-of-turn detection (goal, not yet measured)
Measured1.54 svoice agent reply, caller stops to agent speaks (P6 test calls, turn-gap p50 1.54 s / p95 1.95 s, 13 Nov 2026)
Measured10 Nov 2026same-harness comparison in the table below; human blinded listen pending
Staging4.51predicted MOS (wideband) for the Aria staging path, 10 Nov 2026; not yet on the public API

“Target” figures are goals we have not measured yet, not results. Anything marked “Measured” is a number we took ourselves; your latency depends on network, region and text length. First audio is the time to the first non-empty response body, including the network, not a verified start of playback.

10 November 2026

Same-harness results, one client, one text set.

Apart from the row marked Staging, the seven entries come from an eight-arm parallel run, each with the English voice selected by a pre-declared automatic rule: the highest mean automatic wideband predicted-MOS score among up to three candidates (phone score as a tie-break). The Staging row is a separate run later the same day; it is not available on the public API yet. Latency is client-observed over 129 calls per entry: the median and the 95th percentile. Naturalness figures are automatic MOS estimates, not a human listening test. A human blinded listen is pending.

System First audio, median First audio, p95 Full clip, median Word error Predicted MOS, wide / phone
ElevenLabs (1 of 3)0.145 s0.159 s0.229 s6.51%4.37 / 3.93
VoxVain Aria, staging path2 Staging0.209 s0.283 s0.302 s6.51%4.51 / 4.21
VoxCore (Aria)10.160 s0.226 s0.181 s6.51%4.51 / 4.21
Deepgram0.175 s0.241 s2.820 s6.85%4.43 / 3.91
ElevenLabs (2 of 3)0.202 s0.296 s1.346 s6.51%4.28 / 3.15
xAI TTS0.250 s0.342 s0.805 s5.94%4.28 / 3.69
OpenAI0.552 s0.890 s1.516 s6.85%4.48 / 4.01
ElevenLabs (3 of 3)0.636 s0.760 s2.046 s6.74%4.35 / 3.45

1 VoxCore is the production phone-path service. In this test it returned a complete WAV file, not a stream, and it is not the path the public API serves today. 2 The Staging row is the Aria voice on VoxCore, streamed through a staging copy of the VoxVain API route. It is not available on the public API yet; billing integration and review are in progress. This row is a separate run on 10 November (about 1:20–1:40 PM ET) with the same client, texts, harness and scorers, run on its own rather than alongside the other systems, with the phrase cache off and requests paced about 2.2 s apart; its word error was scored on CPU instead of GPU. The 0 request errors refer to that accepted run of 129 requests; an earlier unpaced attempt hit the per-key rate limit and was discarded. The three ElevenLabs rows are three of their public systems, in first-audio-median order; model and pack names are omitted.

Date: 10 November 2026. Client: one host in US East. Text: the same 43 English lines used on 1 November (4,732 characters, 876 normalized words). Connection: one persistent keep-alive HTTPS session per system; one warmup line per voice was paid and discarded. First audio is time from sending the request to the first non-empty response-body chunk, including the network. Full clip is time to the last byte. Latency n = 129 (three passes of the 43 lines) for each entry, 0 request errors. Quality n = 43; for the parallel-run entries, these were also the voice-selection lines. Predicted MOS is an automatic score, reported on wideband audio and on an 8 kHz phone-channel copy. Word error is an automatic corpus rate after an English text normalizer. Predicted MOS and word error are automatic scores; the method is described on this page. The client machine was under heavy load during the parallel run, which can add noise to every system in it. Systems in this table: three from ElevenLabs; Deepgram; xAI TTS; OpenAI; the VoxCore phone path; and VoxVain Aria through a staging API route. No competitor audio is published here. A human blinded listen of this set is pending.

How to read this. No system leads every column. On this client and text set, an ElevenLabs system had the lowest measured first-audio median (0.145 s, against 0.209 s for Aria through the staging API route); VoxCore had the lowest measured full-clip median (0.181 s); xAI TTS had the lowest measured word-error rate among the entries shown (5.94%); the two VoxVain Aria GPU paths, the VoxCore phone path and the staging path, had the highest rounded predicted-MOS results shown, both 4.51 wideband / 4.21 phone (unrounded, the VoxCore phone path is 4.513 / 4.206 and the staging path 4.512 / 4.207). Those are automatic scores, not a human preference result.

Earlier measurement · 1 November 2026

Aria against five ElevenLabs systems, on localhost.

This is a dated earlier comparison, kept for the record. It used the same 43 lines and automatic scorers, but Aria’s speed was measured on the test node’s localhost as a complete file, not on the public API. It is not the current same-harness result; see the 10 November table above.

Metric VoxVain Aria ElevenLabs A ElevenLabs B ElevenLabs C ElevenLabs D ElevenLabs E
Predicted MOS, wideband4.5124.3474.2804.3604.4374.451
Predicted MOS, phone channel4.2053.8943.2073.4843.9744.102
Word error rate, lighter scorer, phone %6.745.595.946.516.165.94
Word error rate, heavier scorer, phone %7.084.575.945.375.256.16
Entity error %, small / medium0.0 / 1.271.27 / 0.00 / 00 / 00 / 00 / 0
First audio p50 / p90, seconds0.115 / 0.136*0.134 / 0.1480.174 / 0.2320.497 / 0.5570.403 / 0.4640.357 / 0.498

n = 43 lines (4,732 characters). Latency n = 129 (three passes). The ElevenLabs A–E columns are five of their public systems, in the original column order; model, pack and voice names are omitted. An automatic judge on 12 lines, both orders: Aria edged ElevenLabs A (3 wins, 1 loss, 8 ties); ElevenLabs B led Aria (7–1–4); the A/A control was 12 ties. Predicted MOS is an automatic score, not a human MOS. These are automatic scores; the method is described on this page. *Aria latency is a complete, non-streamed file on localhost. Over a network it would be slower; we did not measure that path in this run. An earlier same-harness test from our box, 4 October 2026, had Aria 0.21 s and an ElevenLabs system 0.19 s to first audio.

How to read this. Aria led the automatic naturalness scores, especially on a phone-band copy, and had the lower localhost first-audio median in this set. Every ElevenLabs system had a lower word-error rate. The then-newest of the five in this test, ElevenLabs B, was preferred by the automatic judge. This mixed result is not a human listen and is not a public-API latency test.

List prices, 10 November 2026

Our list prices next to other providers.

Each figure is from the provider's public pricing page, fetched 10 November 2026. Minutes are not packaged identically, so read what each one includes. We are not the lowest in every column.

ProviderSpeech, per 1,000 charactersStreaming transcription, per minuteVoice agent, per minutePhone, per minuteNumber, per month
VoxVain$0.012$0.0052$0.048, conversation model included$0.015 in / $0.019 out$1.59 per additional number
ElevenLabs$0.04$0.0037$0.08n/an/a
Cartesia$0.037 to $0.065$0.0065$0.03 plus the language model$0.014 on a Cartesia numbern/a
Deepgram$0.015 and $0.030$0.0077$0.075n/an/a
OpenAI$0.015$0.017, $0.003 batch$0.05n/an/a
Vapipass-throughpass-through$0.05 hosting plus providerspass-throughn/a
Retellpass-throughincluded$0.07 to $0.31$0.015$2.00
Synthflown/an/aAnnual contract, no public raten/an/a

Sources, fetched 10 November 2026: ElevenLabs, Cartesia, Deepgram, OpenAI, Vapi, Retell, Synthflow; AssemblyAI and Murf figures on Pricing: AssemblyAI, Murf. VoxVain prices are for customers subscribing from 10 November 2026. PlayHT closed on 31 December 2025 and is not listed. Plans and allowances are on Pricing.

How we test

The rules for a number we will stand behind.

Within each test on this page, every system is measured with the same text, the same client, and the same clock. Latency is client-observed time to the first non-empty response body, including the network. Naturalness scores are a predicted MOS from an automatic scorer, not a human rating. The method is described on this page. We do not publish a single-winner claim, and a human blinded listen of the 10 November set is still pending.

This page is not Bench + Sim, which is our agent-testing product. List prices live on Pricing. We do not treat vendor marketing latency, or a public arena ranking, as our measurement.

Start building with VoxVain.

Get a free API key in a minute, no card needed. Then follow the quickstart.