AssemblyAI Launches TTS Purpose-Built for Voice Agents: 90ms Streaming Response

Voice agents are moving from experimental tools into production environments, and latency and cost remain two unavoidable hurdles. AssemblyAI's new product AssemblyAI-TTS-1, launched on Product Hunt, directly targets these two pain points: natural speech synthesis built for voice agents, streaming response as low as 90 milliseconds, and character-based billing instead of annual contracts.

TTS Positioning Built for Voice Agents

Unlike speech synthesis aimed at offline scenarios such as podcasts and audiobooks, voice agents have far stricter real-time requirements. Every millisecond a user waits for a response after finishing a sentence directly affects how natural the conversation feels. AssemblyAI positions TTS-1 as "built for voice agents," with streaming synthesis as its core selling point—audio doesn't wait for the entire text to be processed before output begins; it's pushed out as it's generated.

This positioning also explains why the product emphasizes "unlimited concurrency." The typical load for voice agents involves a large number of short conversations happening simultaneously. Traditional billing by concurrency count or minutes can easily create bottlenecks during peak periods, while unlimited concurrency means integrators don't need to worry about capacity planning.

What 90ms Streaming Response Means

A 90-millisecond streaming latency is the key metric AssemblyAI provides in this launch. In voice conversations, end-to-end latency is typically composed of three parts: speech recognition, language model inference, and speech synthesis. If the TTS stage can be compressed to under a hundred milliseconds, it leaves a more generous budget for the entire pipeline and makes natural conversational behaviors like "interrupt-and-respond" feel closer to human rhythm.

It's worth noting that 90 milliseconds is a streaming first-packet metric, not the generation time for an entire segment of speech. For teams building real-time voice agents, this is precisely the most meaningful number—what users perceive is "how long until I hear the first sound," not "how long it takes for the entire utterance to finish generating."

Time to First Byte (TTFB) and end-to-end latency are two easily confused metrics in voice agent evaluation. First-packet latency refers to the time from when a request is sent to when the first chunk of audio data is received, determining "how long until the user hears the first sound"; end-to-end latency also includes the time for speech recognition to convert user speech to text and for the large language model to generate response text. The industry rule of thumb for voice conversations is: when end-to-end latency is below 500 milliseconds, the conversation experience approaches human-like quality; exceeding 1 second makes it noticeably feel like "the other party is thinking." Therefore, the 90ms first-packet latency in the TTS stage is effectively reserving latency budget for ASR and LLM.

Streaming synthesis (Streaming TTS) works like this: the model doesn't need to wait for the entire text input to be complete before generating audio; instead, it receives text chunks and immediately outputs corresponding audio chunks, pushing as it generates. This requires the model to have low-latency incremental inference capability, and audio encoding and transport protocols (such as WebSocket or HTTP chunked) must work in concert. The significance of streaming synthesis for voice agents is that the LLM's response text is itself generated token by token—if TTS can consume these tokens in a streaming manner synchronously, the latency across the entire pipeline can be "overlapped" rather than accumulated serially.

Pricing: Character-Based Billing, No Annual Contract

On pricing strategy, AssemblyAI offers $15 per million characters with no annual contract required. Character-based billing is relatively friendly for voice agent scenarios because conversation lengths vary widely and usage fluctuates significantly, allowing teams to pay for actual consumption without committing to a fixed quota upfront.

"No annual contract" is also a point emphasized in this launch. For teams still validating voice agent product forms, the room for trial and error with low lock-in cost is often more attractive than the unit price itself.

Character-based billing vs. per-minute/per-concurrency billing: Speech synthesis services commonly use three billing methods—by audio duration (minutes), by concurrent channels, or by input character count. Per-minute billing is intuitive for long audio scenarios (audiobooks) but unfriendly for voice agents dominated by short conversations, because the billing unit is disconnected from the actual text volume consumed; per-concurrency billing requires teams to estimate peak capacity, making it easy to pay for idle capacity. Character-based billing directly corresponds to the actual computational load of the TTS model (input text length), and costs fluctuate with actual usage when volume varies, making it more suitable for agent scenarios with widely varying conversation lengths. At $15 per million characters, that's approximately $0.015 per thousand characters, which can be compared horizontally with the published prices of mainstream TTS providers.

Unlimited concurrency means the server doesn't limit the number of simultaneous synthesis requests, and integrators don't need to pay extra for capacity planning or apply for quotas. This is particularly important for the typical load of voice agents (many short conversations happening simultaneously, with obvious peaks), but it also means actual throughput depends on the provider's elastic scaling capability—stability during peak periods should be a focus when selecting a vendor.

Competition in Voice Agent Infrastructure Is Heating Up

AssemblyAI was previously better known for its speech recognition capabilities. The launch of a dedicated TTS product means it's extending from a single ASR vendor toward complete infrastructure for voice agents. The combination of recognition and synthesis allows developers to complete two key stages of the voice pipeline with the same vendor, reducing latency and engineering complexity from cross-service integration.

Based on Product Hunt data, the product received 70 votes and 2 comments, ranking 20th—a moderate level of popularity, indicating that the market is still in an observation phase regarding the "TTS purpose-built for voice agents" niche. The real test lies in the stability of 90ms latency in real network environments and whether capabilities such as multilingual support and voice diversity can meet the needs of production-grade voice agents.

For teams building voice agents, AssemblyAI-TTS-1 is worth adding to the vendor shortlist for a hands-on test, especially for scenarios that are sensitive to first-packet latency and have highly variable usage.

ASR (Automatic Speech Recognition) and TTS (Text-to-Speech) form the "ears" and "mouth" of voice agents, while the LLM is the "brain." Traditionally, teams often separately procure services from three vendors for ASR, LLM, and TTS, then orchestrate them into a conversation pipeline themselves. Network round-trips from cross-service calls add latency and increase the complexity of troubleshooting and billing reconciliation. AssemblyAI's expansion from ASR to TTS follows a "voice pipeline integration" approach: completing recognition and synthesis within the same vendor can optimize internal pipelines, unify authentication and billing, and reduce engineering integration costs. This trend is becoming increasingly evident in the voice AI space—multiple voice infrastructure vendors are extending from single capabilities toward "full-stack voice agents," with competition shifting from model quality to a comprehensive contest of latency, concurrency, and developer experience.

One-line Summary

AssemblyAI launches TTS-1 purpose-built for voice agents, with 90ms streaming first-packet latency, character-based billing, and no annual contract.

Paragraph Summary

AssemblyAI launched AssemblyAI-TTS-1 on Product Hunt, positioned as a text-to-speech product purpose-built for voice agents, featuring streaming response as low as 90 milliseconds—audio is pushed as it's generated without waiting for the entire text to be processed. The product emphasizes unlimited concurrency, suitable for agent workloads with many simultaneous short conversations; billing is $15 per million characters with no annual contract required, making it easy for teams to experiment based on actual usage. The article notes that 90 milliseconds is a streaming first-packet metric rather than total generation time—precisely the key to users' perception of "how long until I hear the first sound." Meanwhile, AssemblyAI is extending from a single ASR vendor toward integrated recognition-plus-synthesis voice agent infrastructure. With 70 votes and a 20th-place ranking on Product Hunt, the market remains in an observation phase regarding this niche.