TTS Latency 2026: The Production Speed Metric for Voice AI

TTS latency is the gap between when a voice agent finishes generating a response and when the user starts hearing it. Humans respond to each other in about 400 milliseconds. That is not a soft guideline. It is the threshold below which a conversation feels natural and above which something starts to feel wrong, even if you cannot name why. A pause that is slightly too long, a response that takes just a beat too much to arrive: users register it as hesitation, confusion, or incompetence, regardless of how accurate the answer is.

In a voice agent pipeline, TTS is the last model before the user hears a response. By the time the text-to-speech model gets its input, the STT model has already transcribed the user’s speech and the LLM has already generated a response. Whatever time TTS takes comes on top of all of that. Across the models in Coval’s TTS benchmark, the median Time to First Audio is 519ms, well past the 400ms threshold. As of July 23, 2026, Palabra TTS v1 leads Coval’s benchmark across TTS models on latency, with a median TTFA of 103ms. Here is what that number means, how we measure it, and why it matters.

Key Takeaways

• Humans respond in around 400ms. TTS latency is the last thing that eats into that budget before the user hears a response.
• Most latency benchmarks measure throughput, not streaming. Coval measures streaming only, because that is what a call center actually runs on.
• TTFA (Time to First Audio) is the primary TTS latency metric: how long from when the TTS model receives text to when the first audio byte arrives.
• As of July 23, 2026, Palabra TTS v1 leads Coval’s TTS benchmark at 103ms median TTFA. Live rankings at benchmarks.coval.ai.

---

Table of Contents

1. Why Latency Is the Metric Nobody Gets Right
2. Why TTS Latency Matters More Than People Think
3. Throughput vs. Streaming: Why Most Benchmarks Measure the Wrong Thing
4. How Coval Measures TTS Latency: TTFA
5. July 2026 Benchmark Results: Palabra Leads on Latency
6. How to Evaluate TTS Latency for Your Use Case
7. FAQ

---

Why Latency Is the Metric Nobody Gets Right

Ask a vendor for their latency numbers and you will get a figure. Ask two vendors for their latency numbers and you will get two figures measured in completely different ways under completely different conditions.

Throughput, first-byte latency, end-to-end latency, real-time factor: each measures something real, but none of them are the same thing. The number that matters for a production voice agent is specific: how long does a user wait between the agent finishing its reasoning and the agent starting to speak? That number is what Coval benchmarks. Everything else is a proxy.

---

Why TTS Latency Matters More Than People Think

TTS is the last model in the pipeline before the user hears a response. That position matters.

A voice agent pipeline typically looks like this: the user speaks, the STT model returns a transcript, the LLM interprets the intent and generates a response, the TTS model converts that response into audio, and the user hears it. Each of these steps takes time. The 400ms budget for a natural-feeling conversation has to be split across all of them.

TTS latency comes last. By the time the TTS model receives its input, every other step has already consumed part of the budget. A slow TTS model arrives at the end of a chain that may already be close to the threshold, tipping the whole interaction over it.

The inverse is also true. A fast TTS model frees time across the pipeline. When TTS returns audio quickly, the total response time stays inside the 400ms window even when earlier steps have a slower moment. The buffer that keeps conversations feeling natural is largely built by fast TTS.

For a complete breakdown of end-to-end voice pipeline latency (including STT and LLM steps), see Coval’s complete voice AI latency guide.

What low TTS latency buys you: Not just faster audio. More slack in the pipeline for every upstream model.

---

Throughput vs. Streaming: Why Most Benchmarks Measure the Wrong Thing

There are two distinct use cases for text-to-speech: batch generation and real-time streaming.

Batch generation is what you use to produce a completed audio file: a podcast episode, a dubbed video, an audiobook chapter. You send the full text and wait for the full audio file. Speed matters, but the user is not waiting live. According to Artificial Analysis, which measures throughput as a real-time factor (seconds of audio generated per second of compute), this is the right metric for batch work.

Real-time streaming is what a call center voice agent runs on. The LLM has finished generating a response and the agent needs to start speaking as quickly as possible so the conversation does not stall. Throughput tells you almost nothing useful here because the user is not waiting for a file. They are waiting to hear anything at all, and every millisecond before that first audio byte is a millisecond of silence in a live call.

Coval benchmarks streaming latency only, because that is the scenario that matters for voice agents. In Coval’s benchmark, the models that lead on streaming TTFA are not consistently the same models that lead on throughput rankings. The two lists diverge. A model that achieves excellent throughput on batch generation can still feel slow and unresponsive in a live call, and using a throughput benchmark to make a streaming decision is a common and costly mistake.

---

How Coval Measures TTS Latency: TTFA

Since Coval expanded its benchmark coverage from TTS into full STT evaluation, latency has been measured alongside accuracy as a parallel evaluation dimension. For TTS, Coval uses one primary latency metric: TTFA.

Time to First Audio (TTFA)

TTFA measures the latency from when the TTS model receives its text input to when the first audio byte is returned. The timer starts the moment text is sent. It stops when audio begins arriving.

This is the number that directly corresponds to the gap a user experiences between the end of the LLM’s processing and the start of the agent’s voice. Low TTFA means the agent starts speaking quickly. High TTFA means there is a noticeable pause before anything is heard, regardless of how good the audio sounds once it arrives.

Coval reports TTFA at both median and p95. The median tells you typical behavior. The p95 tells you how bad the worst 5% of interactions are. For a call center running thousands of calls a day, a high p95 means a meaningful share of your users are experiencing that delay on every single call.

What TTFA measures: How long between when the agent has something to say and when the user hears the first sound.

---

July 2026 Benchmark Results: Palabra Leads on Latency

As of July 23, 2026, Palabra TTS v1 leads Coval’s TTS benchmark on TTFA. Benchmarks refresh every 30 minutes and the live ranking is always at benchmarks.coval.ai.

Model	Median TTFA
Palabra TTS v1	103ms
TTS-2 (Inworld AI)	196ms
Flash v2.5 (ElevenLabs)	202ms

What the result means: At 103ms median TTFA, Palabra TTS v1 is producing its first audio byte in roughly a quarter of the 400ms human response window. That leaves substantial time for the upstream pipeline (STT and LLM) before the total response time becomes noticeable.

What the gap means: The difference between first and second place is not marginal. Palabra’s 103ms is approximately half the latency of the next closest model. At that delta, users do not need a benchmark to notice. The conversation either feels natural or it does not.

What the result does not mean: Latency and audio quality are separate dimensions. A model that leads on TTFA may not lead on naturalness, accuracy, or voice cloning fidelity. Before making a vendor decision on latency alone, evaluate audio quality under your specific use case conditions. For a full provider comparison, see the Coval TTS provider guide at coval.ai/blog.

---

How to Evaluate TTS Latency for Your Use Case

Step 1: Establish your latency budget. Work backwards from the 400ms target. How much time does your STT model take? How much does the LLM need? What is left for TTS? That number is your TTFA ceiling.

Step 2: Look at p95, not just median. Median performance is what you see in demos. p95 is what your users experience in production. If your call center runs 10,000 calls a day, a p95 of 600ms means 500 users per day are experiencing that delay every call. In Coval’s benchmark, the spread between median and p95 TTFA varies widely across providers. Some models hold close to their median at p95; others diverge by a large margin.

Step 3: Test under your actual conditions. Text length, punctuation patterns, and LLM output style all affect TTFA. A model that is fast on short clean sentences may slow down on the verbose, structured outputs common in enterprise LLM pipelines. Benchmark under conditions that match your deployment.

Step 4: Test latency and quality together. The fastest model is not useful if the audio sounds unnatural or the voice does not match your brand. Both have pass/fail thresholds. For a step-by-step evaluation framework, see how to evaluate voice agents.

Track your chosen model over time with Coval’s Observe. Providers ship silent infrastructure updates that affect latency. A model that leads now may not lead next month. For a framework on embedding latency checks into a continuous evaluation workflow, see Coval’s voice AI regression testing guide.

---

FAQ