Best Low-Latency TTS APIs 2026

By Gradium. Data as of September 2026.

Key takeaways

• This is the latency cluster hub: the whole turn budget, not the synthesis stage alone.
• Since June 3, 2026 the Coval leaderboard reports perceived time to first audio, adding the leading silence inside the stream to the round trip. Any figure published before that date is round trip only.
• On the read of September 8, 2026 (1-day window, 26 models), medians ran from 51 ms for Fluxions vui to 440 ms for Cartesia Sonic 3.6, with ElevenLabs Flash v2.5 at 185 ms and Gradium TTS at 214 ms.
• Read the spread alongside the median. On the same read it ran from 17 ms for the tightest models to 231 ms for Deepgram Aura-2, with Gradium TTS at 31 ms.
• Gradium TTS measures 0 ms leading silence, rank 1 of 27 on the September 10, 2026 board and the only model at zero. Board-wide, leading silence is 28% of the metric.

Latency is the most operationally critical dimension of a TTS API in real-time applications. A 100 ms difference in time to first audio is the difference between a voice agent that feels responsive and one that creates a measurable pause before every response. The 200 to 300 ms threshold above which humans perceive conversational delay sits squarely inside the operating range of every TTS provider on the market, which means the choice of TTS model, and its underlying transport, architecture, and configuration options, directly determines whether your voice agent feels natural or sluggish.

This guide focuses on the latency characteristics of the leading TTS APIs in 2026: what TTFA each achieves, how architecture affects consistency at P99, how configuration options allow per-deployment tuning, and how TTS latency fits into the broader voice agent pipeline. We cover Gradium, ElevenLabs Flash v2.5, Cartesia Sonic 3.6 and Sonic 3.5, and Deepgram Aura-2, drawing on the independent Coval leaderboard read September 8, 2026 and Gradium's own March 2026 round-trip methodology.

What Is TTS Latency and Why Does It Matter?

For the latency figures behind this section, and how Coval has measured perceived time to first audio since June 3, 2026, see How a benchmark change produced a faster TTS model.

Text-to-Speech latency in a real-time context is measured as TTFA (Time to First Audio): the time elapsed between the API receiving the text input and delivering the first audio chunk to the client. It does not measure the time to render the complete audio, which scales with text length. TTFA determines when audio playback can begin.

In batch or file-based TTS use cases (audiobook generation, content dubbing), TTFA is largely irrelevant; total render time and audio quality take priority.

In real-time use cases (voice agents, AI phone calls, interactive assistants), TTFA is the primary latency metric. The human threshold for perceiving a conversational pause as awkward is approximately 200 to 300 ms. A TTS API contributing more than 300 ms TTFA to a pipeline is a structural bottleneck, regardless of how good the underlying audio sounds.

How Does TTS Latency Fit Into the Full Voice Agent Pipeline?

TTS latency does not exist in isolation. A complete voice agent turn involves three latency-contributing stages plus transport overhead:

For conversation to feel natural, the full pipeline should complete in under 800 ms. With STT at 150 ms and LLM at 250 ms, the TTS budget is approximately 200 to 400 ms, depending on network conditions. This is why TTS APIs with TTFA above 300 ms create a structural problem in production voice agents, regardless of audio quality.

With streaming LLMs that emit the first token quickly, the TTS API can begin synthesizing before the full LLM response is available. This LLM to TTS interleaving reduces effective end-to-end latency, but requires the TTS API to support streaming input (receiving and synthesizing text incrementally as tokens arrive). All providers compared in this guide support streaming input; not all use streaming transport.

Why Do P50 and P99 Both Matter?

Published latency benchmarks typically report P50 (median): the latency at or below which 50% of requests complete. P50 represents average-case performance and is the figure providers tend to lead with.

For production systems, P99 (the latency at or below which 99% of requests complete) is the operationally relevant metric. P99 determines the worst 1% of user experiences, which in high-volume deployments represents thousands of interactions per day.

An API with a 90 ms P50 but a 600 ms P99 produces frequent conversation-breaking pauses in production. An API with a 250 ms P50 and a 280 ms P99 delivers highly consistent performance. For voice agents, P99 consistency often matters more than headline P50 figures, and tail latency is where architectural choices (transformer vs State Space Model, batched vs streaming inference) show up most clearly.

How Does WebSocket Streaming Affect TTS Latency?

Beyond model inference time, the transport layer contributes latency on every request.

• HTTP chunk transfer encoding opens a new TCP connection per request. The connection handshake (TCP + TLS) adds approximately 40 to 100 ms per turn, accumulated across every exchange in a conversation. Across a multi-turn conversation that cost is paid on every turn, and it does not appear in published benchmarks: the Coval board excludes the handshake uniformly, roughly 50 to 200 ms, precisely so the comparison is of synthesis rather than of network position.
• WebSocket-based streaming maintains a persistent bidirectional connection. After the initial handshake, subsequent turns incur no connection overhead. For multi-turn voice agents, WebSocket architecture is a significant latency advantage that does not show up in single-request benchmarks.
• WebSocket multiplexing takes this further: reusing a single connection concurrently across multiple streams, further reducing per-turn overhead. Gradium supports WebSocket multiplexing, reducing its effective TTFA to 214 ms in multiplexed configurations (vs 228 ms in standard 32-codebook configuration).

ElevenLabs Flash v2.5

ElevenLabs Flash v2.5 was explicitly designed for real-time conversational AI use cases, as an alternative to ElevenLabs Multilingual v2, which targets audio quality over latency and is not on the current Coval board (benchmarks.coval.ai/tts). On the Coval leaderboard read September 8, 2026, Flash v2.5 recorded 185 ms median perceived time to first audio.

Latency Profile

• Perceived time to first audio: 185 ms median, 25 ms P25 to P75 spread (Coval, 1-day window, read September 8, 2026)
• Transport: HTTP chunk transfer encoding (not a persistent WebSocket)
• P99: not publicly published

What Is the Latency Tradeoff?

On the Coval read of September 8, 2026, ElevenLabs Flash v2.5 recorded 185 ms median perceived time to first audio, Gradium TTS 214 ms and Cartesia Sonic 3.6 440 ms. In a pipeline where Speech-to-Text contributes 150 ms and LLM first token contributes 250 ms, the difference between 185 ms and 288 ms TTS TTFA is 133 ms of total pipeline latency, enough to push a borderline pipeline over the 800 ms naturalness threshold.

Transport matters on top of the model. A connection opened per turn pays a TLS and WebSocket handshake first, roughly 50 to 200 ms depending on network position, and the Coval board excludes that cost uniformly for every model. Gradium's own March 2026 measurement put the difference between per-turn and reused connections at about 44 ms at the median.

Languages and Pricing

A substantially longer language list than Gradium's five; check the ElevenLabs models page for the current count. Flash v2.5 and Eleven v3 Conversational list at $50 per 1M characters as of September 2026, and ElevenLabs ran a 50%-off-for-life API promotion through September 11, 2026. The effective per-character cost depends on the subscription tier.

Best For

ElevenLabs Flash v2.5 is the right choice when ElevenLabs' voice catalogue is a hard requirement, broad language coverage matters, and the use case is low to medium turn-frequency (where per-request connection overhead is less material). For a deeper Gradium-vs-ElevenLabs breakdown, including TTS, STT, voice cloning, and deployment, see the ElevenLabs alternative comparison.

Cartesia Sonic 3.6

Cartesia built its Sonic line on State Space Model (SSM) architecture, and claims lower time to first audio, a lower real-time factor and higher throughput from it. Note what that claim does not cover: on the Coval read of September 8, 2026, Sonic 3.6 recorded a 171 ms P25 to P75 spread, the second widest on the board, so the architecture claim is about throughput rather than about variance.

Latency Profile

• Perceived time to first audio: 440 ms median for Sonic 3.6, 269 ms for Sonic 3.5 (Coval, 1-day window, read September 8, 2026)
• Transport: WebSocket and REST (both available)
• Spread: 380 ms P25 and 551 ms P75 on the Coval read of September 8, 2026, a 171 ms range, against 31 ms for Gradium TTS and 25 ms for ElevenLabs Flash v2.5 in the same read.

Why does the spread matter?

Any Text-to-Speech model can spike under load or on awkward input: long sentences, unusual phoneme sequences, dense punctuation. In deployments running thousands of concurrent sessions, the spread rather than the median determines how often a conversation-breaking pause happens.

What Cartesia claims from its State Space Model architecture is lower time to first audio, a lower real-time factor and higher throughput, not lower variance. The September 8, 2026 board is consistent with that: Sonic 3.6's 171 ms P25 to P75 spread was the second widest of the 13 models with a reportable median, behind Deepgram Aura-2 at 231 ms.

Languages and Pricing

A substantially longer language list than Gradium's five, with regional accent variants; check the Cartesia models page for the current count. Pricing is credit-based rather than per character: Free at $0 for 20k credits, Pro at $5 for 100k, Startup at $49 for 1.25M and Scale at $299 for 8M, read September 8, 2026.

Best For

Cartesia Sonic 3.6 is the right choice when language coverage or voice preference is the primary requirement: it ranked first of 92 at 1,282 Elo on the Artificial Analysis provider-voice board read September 8, 2026. On the Coval latency read of the same date it was the slowest model in this comparison, so it is not the pick when tail latency impacts a meaningful share of user interactions. For a side-by-side Gradium-vs-Cartesia breakdown, see the Cartesia alternative comparison.

Gradium

Gradium was founded by Neil Zeghidour, Laurent Mazaré, Olivier Teboul, and Alexandre Défossez, co-founders of Kyutai. Kyutai released Moshi (real-time Speech-To-Speech) and Hibiki (live Speech-To-Speech translation). Gradium's latency profile is distinct from the other providers in this comparison because it exposes configurable precision to latency tradeoffs through its codebook architecture, derived from Kyutai's Delayed Streams Modeling research.

Latency Profile

On the independent Coval leaderboard, 1-day window read September 8, 2026, Gradium TTS records 214 ms median perceived time to first audio with a 31 ms P25 to P75 spread, and 0 ms leading silence, rank 1 of 27 on that metric on the September 10, 2026 read. On Gradium's own March 2026 round-trip methodology (Paris, 15 to 25 word sentence, WebSocket, 100 queries, warm), Gradium measures P50 258 ms and P95 274 ms end-to-end, or P50 214 ms and P95 228 ms excluding connection establishment.

Codebook depth is the architecture-level tradeoff behind that latency. More codebooks give higher audio resolution and a higher time to first audio; fewer give the reverse. Gradium tunes it when it trains and ships a model, so it is not a setting exposed in json_config. The measured shape of the tradeoff:

Source: Gradium, "Optimizing quality vs latency in real-time TTS AI models", February 11, 2026. Measured on an NVIDIA RTX 4080 Super at batch 8. Codebook depth is tuned by Gradium at the architecture level; it is not a json_config parameter.

The settings that are exposed per session are temp, cfg_coef, padding_bonus and rewrite_rules, covered in the json_config guide.

WebSocket Multiplexing

Gradium's TTS API is WebSocket-native. Connection multiplexing reuses a single persistent connection across multiple conversation turns, reducing effective TTFA to 214 ms in standard configuration by eliminating per-turn connection overhead.

Gradium measured that saving at about 44 ms at the median in March 2026: 258 ms round trip opening a connection per turn against 214 ms on a reused one. It is paid back on every turn of a conversation, and it is invisible in single-request benchmarks because Coval excludes the handshake uniformly across every model.

CUDA Graph Optimization

Gradium's inference stack uses CUDA Graph optimization, which reduces model inference overhead on NVIDIA GPUs (L4, A10, H100). This contributes to low variance between P50 and P90 latency figures, and is part of why Gradium's published end-to-end P95 (274 ms) sits only 16 ms above its P50.

Total Pipeline Latency

With a streaming LLM (GPT-4 Turbo, Claude), Gradium's full pipeline (STT + LLM + TTS) achieves 420 to 520 ms total turn latency, compared to 2.5 to 5.5 seconds with non-streaming architectures. The WebSocket-native design enables LLM to TTS interleaving: TTS synthesis begins as the LLM emits its first tokens, before the full response is available. For an end-to-end voice agent walkthrough, see how to build a voice AI agent with Gradium and LiveKit.

Languages

English, French, Spanish, German, Portuguese, with regular updates. Mid-sentence code-switching is supported across all five languages with no latency penalty.

Pricing

Credit-based plans starting at $0/month for the free tier (45,000 credits, roughly 1 hour of TTS) and scaling to $1,615/month on Plan L. See Gradium pricing for the full plan breakdown.

Best For

Gradium offers the best balance between configurable low latency, WebSocket-native streaming, and unified TTS + STT infrastructure. Teams that need to tune latency vs audio quality per deployment (different codebook configs for different endpoints) benefit from Gradium's precision controls. The multiplexing advantage is particularly valuable in high-turn-frequency voice agent deployments. For a deeper dive on Gradium for voice agents specifically, see the best Text-to-Speech API for voice agents.

Deepgram Aura-2

Deepgram Aura-2 is a low-latency TTS model integrated into Deepgram's STT-first platform. It is designed to pair with Deepgram Nova-3 in voice agent pipelines.

Latency Profile

• Perceived time to first audio: 290 ms median, 231 ms P25 to P75 spread (Coval, 1-day window, read September 8, 2026)
• Transport: WebSocket
• Languages: 7 (English, Spanish, French, German, Dutch, Italian, Japanese)

Deepgram's Voice Agent API uses Aura-2 as the TTS component within a bundled STT + LLM + TTS orchestration endpoint, reducing integration overhead for teams already using Deepgram Nova.

Pricing

$0.030 per 1,000 characters ($0.027 at Growth tier).

Best For

Deepgram Aura-2 is the right choice for teams already using Deepgram Nova-3 for STT who want to consolidate latency-critical infrastructure on a single vendor. HIPAA-compliant on-premise deployment is available. For a side-by-side Gradium-vs-Deepgram breakdown, including TTS, voice cloning, and deployment, see the Deepgram alternative comparison.

How Should You Match TTS Latency to Your Use Case?

Not all real-time applications have the same latency tolerance. The table below maps TTFA ranges to application requirements.

How Should You Choose a Low-Latency TTS API?

• Choose ElevenLabs Flash v2.5 if ElevenLabs' voice quality and broad language catalogue (32 languages) are primary criteria alongside competitive latency. Plan for HTTP streaming overhead in multi-turn deployments.
• Choose Cartesia Sonic 3.6 if voice preference or broad language coverage is the requirement. It led the Artificial Analysis provider-voice board on September 8, 2026 at 1,282 Elo, and was the slowest model in this comparison on the Coval latency read of the same date.
• Choose Gradium if you need a low median with a tight spread (214 ms and 31 ms on the Coval read of September 8, 2026) and no leading silence at all, WebSocket-native streaming with multiplexing, unified Text-to-Speech and Speech-to-Text infrastructure, and the ability to tune precision vs latency per deployment context. The full pipeline latency of 420 to 520 ms (STT + LLM + TTS) with streaming LLMs is among the lowest in production voice agent stacks.
• Choose Deepgram Aura-2 if you are already on Deepgram Nova for STT and want consistent low-latency TTS without adding a new vendor.

Also comparing Cartesia, ElevenLabs, or Deepgram head-to-head with Gradium? Each comparison goes deeper on TTS quality, STT, voice cloning, and deployment options.

Glossary

Perceived time to first audio. Round-trip time to first audio plus the leading silence inside the stream before the first audible sample. The Coval headline metric since June 3, 2026.

Leading silence. Inaudible samples at the start of a synthesized stream, measured as everything before the first 10 ms window whose RMS exceeds 0.01 on a normalized signal.

P25 to P75 spread. The gap between the 25th and 75th percentile of latency, also called the interquartile range. A direct measure of consistency that the median cannot show.

Real-time factor (RTF). How many seconds of audio a model generates per second of compute. Higher means more headroom at concurrency.

WebSocket multiplexing. Carrying several synthesis sessions over one persistent connection, so no turn pays the handshake cost.

Turn budget. The total latency a conversational turn may consume across Speech-to-Text, the LLM and Text-to-Speech before it stops feeling conversational.

Related guides

This page is the hub of the latency and streaming cluster. Its siblings:

• Streaming TTS audio in real time: bidirectional streaming and how to measure it correctly.
• WebSocket multiplexing on Gradium TTS: reusing one connection across turns, the cheapest latency win available.
• Connecting Text-to-Speech to an LLM pipeline: interleaving the LLM and the synthesis stage.
• TTS latency benchmark 2026: how perceived time to first audio is measured, and what changed in June 2026.
• json_config for Gradium TTS and STT: the per-session parameters that shape output.
• Turn-taking and VAD in voice agents: the stage upstream of synthesis, where most perceived delay actually lives.

Beyond this topic

Latency is only worth optimizing if the words come out right. TTS WER benchmark 2026 covers accuracy, and making Text-to-Speech pronounce numbers, dates and phone numbers covers the content a live agent reads most.

Gradium's free tier includes 45,000 credits per month with no credit card. API reference and quickstarts are at docs.gradium.ai. For enterprise evaluations, use our contact form.