betaOpen benchmark · refreshed every run

The independent voice AI benchmark for Latin America

We measure latency, error and price of text-to-speech and speech-to-text models with real human audio for every accent. No paid votes, no made-up numbers.

See results How we measure

Last runOct 16, 2026

#1 STT

Neutral (LatAm)

whisper-large-v3-turbo

2.7%

WER

#1 TTS

Neutral (LatAm)

kokoro-em_alex

1.0%

WER

models measured

accents

human clips

US$0

cost per run

Why it exists

Global benchmarks listen in English.

A voice agent that works in New York can fail in Córdoba, Medellín or Recife. We measure what matters for real customers in the region: each accent separately, with human audio under open licenses.

Rioplatense

Voseo and the Buenos Aires / Montevideo sound.

Human reference

Mexican

The largest Spanish-speaking market in the world.

STT · Coming soon

Colombian

Andean accent, huge contact-center volume.

Human reference

Neutral (LatAm)

General Latin American Spanish for multi-country apps.

Human reference

Portuguese (BR)

Brazil: the region's largest voice market.

Human reference

Leaderboard

Speech to text (STT) / Text to speech (TTS)

Human reference

“¿Quieres que te mande el artículo?”

1
whisper-small
faster-whisper · 20 samples
open source
WER
7.5%
p50
1.05 s
CER
3.3%
2
whisper-large-v3-turbo
faster-whisper · 20 samples
open source
WER
22.6%
p50
3.82 s
CER
15.5%
3
whisper-base
faster-whisper · 20 samples
open source
WER
26.3%
p50
480 ms
CER
16.5%
4
whisper-tiny
faster-whisper · 20 samples
open source
WER
27.4%
p50
372 ms
CER
10.7%

#	Model	Error (WER) ↓	Latency p50 ↓	CER	Price
1	whisper-small faster-whisper · int8 · CPU · 20 samples	7.5%	1.05 s	3.3%	Free · open source
2	whisper-large-v3-turbo faster-whisper · int8 · CPU · 20 samples	22.6%	3.82 s	15.5%	Free · open source
3	whisper-base faster-whisper · int8 · CPU · 20 samples	26.3%	480 ms	16.5%	Free · open source
4	whisper-tiny faster-whisper · int8 · CPU · 20 samples	27.4%	372 ms	10.7%	Free · open source

Coming soon · STT

AssemblyAI Universal

From US$0.15 / hour

Coming soon

Deepgram Nova-3

US$0.0077 / min (streaming)

Coming soon

Speechmatics

From ~US$0.24 / hour

Coming soon

Gladia

US$0.61 / hour (async)

Coming soon

Google Cloud STT (Chirp)

US$0.016 / min (streaming)

Coming soon

OpenAI gpt-4o-mini-transcribe

US$0.003 / min

Coming soon

Commercial providers: we will measure them once we can do so with our own accounts and within their terms. Until then we show only their official price, never estimated metrics. Latency = total time to transcribe one sentence (CPU).

Voice agent platforms

Public pricing and Latin America focus, with a source for every data point. Always confirm on the official site.

Retell AI

Low-codeGlobal

From US$0.07/min with models included in the base plan

Global platform; n8n, HubSpot and Twilio integrations

Vapi

DeveloperGlobal

US$0.05/min platform fee + models; real cost ~US$0.13–0.33/min

100+ languages; you pick every model

Bland AI

Global

US$0.09/min includes LLM, STT and TTS; telephony billed separately

Global platform

Synthflow

Enterprise

Sales-led (no self-serve signup)

Enterprise focus

Fonema

LatAmWhatsApp

Pay as you go; telephony US$0.0152/min (Mexico), US$0.07/min (Colombia)

Built for Latin American Spanish: 200+ regional voices, latency < 1.2 s, voice + WhatsApp

Kosmo

Mexico

Plans from MXN 1,290/month

Bills in Mexican pesos

Voice agents →

Reach the people choosing the region's voice infrastructure.

Product teams, agencies and contact centers comparing vendors before they buy.

See media kit →

The weekly pulse of voice AI in LatAm

Ranking changes, new models and prices. One email a week, no spam.