AssemblyAI Ships AssemblyAI-TTS-1: A $15 per 1M Character Text-to-Speech API for Voice Agents

AssemblyAI just dropped its first text‑to‑speech model, AssemblyAI‑TTS‑1, and the headline is the $15 per million‑character price tag. That’s cheap enough to make a VC‑funded hype‑machine look like a bargain basement, but the devil is in the details – latency, feature gaps and lock‑in to a single cloud. As a founder who has been wiring up live voice agents for a decade, I’m looking at how this actually plays out for independent hosting outfits that can’t afford to sit on a hyperscaler’s pricing sheet.

Pricing: The low‑ball headline vs the hidden costs

The $15 per 1 M characters price is indeed the lowest‑priced frontier TTS model that AssemblyAI is marketing for voice agents. It covers every voice, instant cloning, word timestamps and EU processing with no extra per‑voice or concurrency fees. That sounds like a sweet deal for a startup that wants to keep its cost base flat while scaling call volume.

But the competition isn’t sleeping. Inworld’s TTS‑2 Flash also lists $15 per 1 M characters on demand and drops to $7 under a paid plan. The real edge is concurrency: AssemblyAI offers unlimited concurrency on a pay‑as‑you‑go basis, whereas Inworld caps at five concurrent requests in its on‑demand tier. For a boutique provider that needs to burst traffic during peak call windows, that unlimited concurrency could be a lifesaver – provided you can tolerate the latency numbers we’ll get into next.

Latency: Numbers that matter on a live call

AssemblyAI touts a 47 ms median time‑to‑first‑byte (TTFB) in its own test, which is impressive on paper. In the real‑world Cekura benchmark, however, the median streaming latency – measured from the last word sent to the first audible sound – sits at 159 ms. That’s a noticeable lag when you’re trying to keep a caller on the line; every extra hundred milliseconds feels like a dead‑air gap that can make a prospect hang up.

By contrast, Inworld’s TTS‑2 Flash posted a 45 ms median streaming latency in the same independent test, beating AssemblyAI by a wide margin. If you’re building a high‑volume outbound dialer or a real‑time support bot, that latency delta could translate into higher conversion or lower abandonment rates. The takeaway? AssemblyAI’s latency claim is respectable but not market‑leading, and you need to benchmark it against your own call flow to see if the trade‑off is worth the price simplicity.

Feature gaps: The missing SSML and pronunciation tools

At launch, AssemblyAI‑TTS‑1 comes without SSML, speed control, pitch adjustment, emotion tags or custom pronunciation dictionaries. The WebSocket API can’t accept any of those parameters, and the HTTP route is similarly bare. For developers who need nuanced voice control – think dynamic emphasis, brand‑specific pronunciation of product names, or regional dialect tweaks – you’ll hit a wall.

Competitors are already offering those bells and whistles. Cartesia’s Sonic 3.6 includes custom pronunciation dictionaries, and ElevenLabs’ Flash models support SSML. If your product roadmap includes multi‑tone voiceovers or brand‑specific vocal styles, you’ll have to build work‑arounds or wait for AssemblyAI’s “coming soon” roadmap items. That’s a risk for any founder who can’t afford to postpone a feature launch.

Voice cloning: Instant but limited

The launch includes a library of over 4,000 voices and instant voice cloning that works with a 24‑second audio clip, provided you have speaker permission. That’s handy for quickly rolling out a personalized agent voice without a long fine‑tuning process. However, the instant clone is the only offering; professional, fine‑tuned clones aren’t available yet.

For a startup that wants a unique brand voice, the instant clone is a good start, but you’ll eventually need a more robust solution if you want high fidelity or a larger voice catalog. Inworld already offers both instant and professional cloning, giving them a broader runway for voice‑centric products.

Deployment constraints: Cloud‑only and regional limits

AssemblyAI‑TTS‑1 runs only in their managed cloud – US East (us‑east‑1) or EU (eu‑north‑1). There’s no on‑premise or VPC‑private option, which means you’re handing over your audio data to a third‑party provider. For regulated industries or customers with strict data residency requirements, that could be a deal‑breaker.

On the flip side, the hosted model means you don’t have to worry about scaling GPU clusters or managing model updates. If your business model is purely SaaS and you’re comfortable with the US/EU data flow, the simplicity can outweigh the lack of on‑prem flexibility.

Cost modelling: What $15 per million characters really means

AssemblyAI’s pricing calculator assumes 600 characters per minute of phone speech. At that rate, a 10,000‑minute month translates to roughly 6 M characters, costing about $90 per month. That’s a modest baseline for a small‑to‑mid‑scale operation. The unlimited concurrency means you can spike traffic without incurring extra fees, which is a big plus over tiered‑concurrency models that force you to buy higher‑priced plans for burst capacity.

However, remember that the price is per character, not per minute. If your prompts are verbose or you embed a lot of filler text, costs can climb quickly. Also, the lack of per‑voice fees is nice, but you’ll still pay the same $15 for every cloned voice you generate, so a large catalog of custom voices could add up.

Strategic takeaways for independent hosting providers

AssemblyAI‑TTS‑1 is a solid entry point for founders who need a low‑cost, unlimited‑concurrency TTS engine to close the loop on speech‑to‑text pipelines. Its pricing simplicity and instant cloning are strong draws, especially for early‑stage products that need to move fast.

But the latency lag behind Inworld, the missing SSML and pronunciation controls, and the cloud‑only deployment model are real friction points. If your product hinges on fine‑grained voice control, low latency, or strict data residency, you’ll need to weigh those gaps against the price advantage. My advice: run a pilot with real call traffic, measure abandonment rates at the 150‑ms latency threshold, and decide if the trade‑off is acceptable. Keep an eye on AssemblyAI’s roadmap – they’ve flagged the missing features as “coming soon,” but until they land, you may have to supplement with a secondary TTS vendor or build custom post‑processing.

In the end, the market is still wide open for a nimble provider who can stitch together the best of both worlds – the cheap, unlimited‑concurrency model of AssemblyAI and the feature depth of the bigger players. That’s where the real opportunity lies for anyone willing to roll up their sleeves and engineer a hybrid solution.

— Allan Ali, Founder

This article was produced with AI-assisted research and editorial support. Reporting is based on the source material cited below. Sources: MarkTechPost; marktechpost.com; Global1.News (30 October 2026).

By Allan Ali, Global1.News