Developers

Build on VoxVain.

One HTTPS API for speech, transcription, turn-taking, agents and phone. Here is what runs underneath, what is available today, how it measures, and how to make your first call.

Live comparison. Measured cells come from the probe feed. Price and feature cells come from dated official pages.
SpeedQualityReliabilityPriceFeatures
Namep50p95p99s50s95s99ScoreWERCERRTF24h successErrnAge$/1M$/minVoicesLangsStreamCharsFreeLimitRegion
VoxVain———————————————————————
ElevenLabs———————————————————————
OpenAI———————————————————————
Deepgram———————————————————————
Murf———————————————————————
Cartesia———————————————————————
Vapi———————————————————————
Retell———————————————————————
Synthflow———————————————————————

Ranges show published tiers. Checked 12 Nov 2026 · VoxVain list price · VoxVain streaming · ElevenLabs · OpenAI · Deepgram · Murf. Cartesia, Vapi, Retell, and Synthflow were not re-checked.

Where we lead

Measured results, with the date and the method.

Measured4.51Highest rounded predicted MOS, wideband, shared by our two Aria GPU paths: the VoxCore phone path in the parallel run (4.513) and Aria on a staging path in a separate run (4.512). Next: OpenAI 4.48, Deepgram 4.43, ElevenLabs 4.37. 10 Nov 2026. Method
Measured4.21Highest rounded predicted MOS on a phone-band copy (8 kHz), shared by the same two paths: VoxCore phone path 4.206, Aria staging path 4.207. Next: OpenAI 4.01, ElevenLabs 3.93. 10 Nov 2026. Method
Measured0.181 sLowest full-clip median in the parallel run, for the VoxCore phone path, which returns a complete file and is not the public API. Next: ElevenLabs 0.229 s. 10 Nov 2026. Method
Measured4.512Aria’s wideband predicted MOS against five ElevenLabs systems in our earlier test; the highest of those was 4.451. Every ElevenLabs system had a lower word-error rate that day. 1 Nov 2026. Method
Measured4.205Aria’s phone-band predicted MOS in the same earlier test; the highest ElevenLabs system was 4.102. Every ElevenLabs system had a lower word-error rate that day. 1 Nov 2026. Method

Honest comparison

Where others are ahead.

Others ahead0.145 sAn ElevenLabs system had the lowest first-audio median, against 0.160 s VoxCore TTS first audio (production phone path) and 0.209 s for Aria on a staging API route. 10 Nov 2026. Method
Others ahead5.94%xAI TTS had the lowest word-error rate among the entries shown. VoxVain Aria was 6.51%, tied with two of the ElevenLabs systems. 10 Nov 2026. Method

Automatic scores from dated same-harness tests: latency over 129 calls per entry, quality over 43 lines, each provider shown with the best of its tested candidate voices. Predicted MOS is an automatic score, not a human rating. Each figure links to its method. Scoring tools are listed on the open-source acknowledgements page. Results as of 10 Nov 2026. Full tables and method on Compare.

What’s available

Status, in one table.

ProductStatusWhat it does
Voice APIPreviewText to speech in 21 voices across 12 languages, as a file or a stream.
Aria, staging pathStagingAria on VoxCore through a staging API route. Not on the public API yet; billing integration and review are in progress.
Listen and TurnPreviewStreaming and batch speech to text, with turn detection.
Voice Agents and EnginePreviewReal-time agents: listening, turn detection, the conversation engine and speech in one minute.
PhoneLivePhone numbers and minutes for agents that answer real calls.
CloneComing soonA consented custom voice, used through the same API.
Pro StudioTodayThe browser voice booth: try voices, tune and export takes.

Status labels match Pricing and Status.

Technology

VoxCore: real-time voice, measured.

VoxCore

Our real-time voice platform. VoxCore TTS first audio, median 0.16 s (10 Nov 2026, 129 calls, US East, production phone path). The full clip arrived in 0.18 s across 129 calls with 0 errors. On a GPU it renders speech more than 100× faster than real time in under 1 GB of GPU memory (10 Nov 2026).

GPU serving

Live Talk and the Aria staging path run on VoxCore. In a staging test, a request sent while the GPU worker was drained was served by the existing CPU voice instead of failing.

Streaming API

Plain HTTPS: /v1/text-to-speech/{voice} for a file and /stream for chunks, sentence by sentence. Voices and usage have their own endpoints in the API reference.

Audio formats

MP3, WAV, raw 16-bit PCM at 24 kHz, and 8 kHz μ-law for phone lines.

Clone building

A short recording plus a spoken consent phrase with one-time words. The recording is checked with speech recognition and a speaker-embedding check before the voice is built.

Delivery quality

VoxCore trims, fades, and normalizes loudness to the delivery target, and can convert audio for phone lines. On 10 Nov 2026, all 858 target/clip cases (286 clips across three delivery targets) were byte-identical across the three versions tested.

Workflow

Build on VoxVain in four steps.

1. Get a key

Sign up free, no card needed, and create an API key. Get API key

2. Pick or clone a voice

Choose from 21 voices, try them in the booth, or make a consented Clone.

3. Integrate

Call the HTTPS API from your app, or put an agent on a phone number with Phone.

Quickstart

Your first audio.

Set VOXVAIN_API_KEY in your environment first; the examples never contain a real key. Python needs requests (pip install requests); the JavaScript runs on Node 18 or later, saved as an .mjs file.

Text to speechPreview API
curl -X POST https://api.voxvain.com/v1/text-to-speech/aria \
  -H "Authorization: Bearer $VOXVAIN_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"text": "Hello from VoxVain.", "output_format": "mp3"}' \
  --output hello.mp3
Streaming, phone-band audioPreview API
curl -N -X POST https://api.voxvain.com/v1/text-to-speech/aria/stream \
  -H "Authorization: Bearer $VOXVAIN_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"text": "Thanks for calling.", "output_format": "mulaw_8000"}' \
  --output reply.ulaw