How Speechify’s Simba 3.2 Took the No.1 Spot on Voice AI’s Toughest Independent Benchmark

Jul 11, 2026 02:00 AM - 7 hours ago 48

Opinions expressed by Entrepreneur contributors are their own.

For years, the norm successful text-to-speech has been simple. If you wanted the best-sounding sound for your product, you paid endeavor pricing. If you wanted cheap, you accepted robotic. If you wanted fast, you gave up thing connected both. That norm conscionable broke.

The trade-off each merchandise squad has been forced to make

If you person ever built a sound agent, a telephone system, aliases a real-time reader, you cognize the drill. You audition 4 aliases 5 models. One sounds unthinkable and costs much than your infrastructure. One is affordable and sounds for illustration a GPS from 2009. One is fast, but only successful 3 languages. You prime the slightest bad action and ship.

Then the invoice arrives.

And each quarter, your CFO asks the aforesaid question: why is sound the azygous astir costly statement point successful the stack?

What conscionable changed connected the leaderboards

This week, Speechify’s Simba 3.2 moved to first spot connected the Artificial Analysis text-to-speech leaderboard, ranking supra ElevenLabs, Cartesia, OpenAI, and Google DeepMind. On Voice Arena, the blind-listener benchmark modeled connected Chatbot Arena, it sits astatine the apical for real-time models astatine its value point.

Neither leaderboard is tally by Speechify. Neither uses self-reported scores. Native speakers perceive 2 clips without knowing which exemplary made which, and they ballot for whichever sounds much natural.

Simba 3.2 is now the highest-rated real-time sound exemplary a squad tin put successful accumulation today.

Here is wherever it gets uncomfortable for the incumbents.

The 3 numbers that matter

For anyone building pinch voice, only 3 things ever really mattered: quality, latency, and cost. Every exemplary merchandise has forced a discuss connected astatine slightest 1 of them.

1. Quality. Simba 3.2 is classed number 1 connected Artificial Analysis and connected apical for value and value connected Voice Arena. Both benchmarks are independent. Both are blind.

2. Latency. It is simply a streaming-native exemplary pinch little time-to-first-byte than its predecessors, built for sound agents that respond successful existent clip alternatively than aft a region that ruins the conversation. All sub-100ms.

3. Cost. It is listed astatine $10 per 1 cardinal characters, dropping to $6 per 1 cardinal characters connected the Scale tier. That makes it the cheapest exemplary successful the Artificial Analysis apical ten, complete 15 times much affordable than ElevenLabs and astir six times much affordable than Cartesia, according to the company.

Best-sounding, fastest, and cheapest person almost ne'er described the aforesaid model. Now they do.

Credit: Speechify

Why this happened

The accustomed communicative pinch AI models is that the laboratory optimizes for the benchmark, prices for endeavor buyers, and lets the developer level inherit immoderate separator is near over. Speechify built it successful the other order.

The aforesaid sound exertion has been moving wrong a user merchandise utilized by much than sixty cardinal group for years. That assemblage does not tolerate a robotic voice, a two-second hold earlier the first word, aliases the benignant of portion economics that only activity astatine endeavor pricing. Every A/B trial successful that merchandise fed backmost into the model.

“We made the architecture decisions astatine the opening that astir labs put disconnected until later,” explained Raheel Kazi, an engineering leader astatine Speechify. “We ne'er wanted to sacrifice connected costs to pursuit quality, aliases sacrifice connected value to pursuit latency. We took the harder way connected purpose. Hitting SOTA connected each 3 astatine erstwhile is what that determination was ever for.”

“This is the underdog communicative for API providers,” Luke Oliff, Head of Developer Relations astatine Speechify, said successful a press release. “We spent years making our models tally efficiently because our user business demanded it, tens of millions of listeners, pinch immoderate of the champion voices connected the planet. That activity is why we tin now put the best-rated exemplary successful the world connected our API astatine astir arsenic inexpensive arsenic it comes. Most labs are built for the benchmark and priced for the enterprise. We built for listeners and priced for production.”

What Artificial Analysis and Voice Arena really test

Neither leaderboard is the benignant of benchmark a vendor tin game.

Artificial Analysis runs connected unrecorded serverless API endpoints, 4 times a time astatine random times, utilizing a randomly selected voice, a unsocial 500-character prompt, and a standardized audio sample rate. Latency is measured end-to-end, each the measurement to erstwhile the audio record lands locally.

Voice Arena uses the aforesaid unsighted pair-comparison rule crossed six languages, pinch a balanced sound slate per exemplary alternatively than each vendor’s best-sounding default. The methodology was developed pinch input from Prof. Shinji Watanabe of Carnegie Mellon University.

On some boards, value is scored the aforesaid way. Pairs of clips generated from identical matter are played to autochthonal speakers successful unsighted comparisons. Listeners take which sounds much natural. Votes get aggregated into an Elo rating. No self-reported score, nary vendor-selected clip, nary soul panel, and nary supplier pays for inclusion aliases ranking.

For a exemplary to beryllium adjacent the apical of both, it has to fulfill an nonsubjective capacity information and a unsighted quality penchant ballot crossed aggregate languages. Simba 3.2 does.

SpeechifyAI Agents and Speechify’s Developer Platform

Alongside the leaderboard result, Speechify is launching Voice Agents for businesses and a developer platform, some astatine speechify.ai. The exemplary powering some is the aforesaid 1 moving its user apps.

Simba 3.2 is simply a streaming-native exemplary pinch debased time-to-first-byte, fine-grained affectional control, and SSML prosody, engineered to sound earthy successful real-time sound applications. According to the company, much voices, further languages, and an moreover lower-cost tier are already connected the roadmap.

“Simba 3.2 is our champion exemplary yet, now disposable on Speechify.ai,” Cliff Weitzman, CEO and Founder of Speechify, shared successful a public post. “It’s built to powerfulness sound agents astatine standard and perfected from millions of A/B tests we tally successful our user platform. In TTS APIs, 3 things matter: cost, quality, and latency. Simba 3.2 has achieved SOTA connected this trifecta. Beyond excited for you to acquisition it firsthand to powerfulness your experiences.”