Data read October 22, 2026 · Coval, Speko, Artificial Analysis · measured from the US East coast, in English
Low-latency TTS for voice agents: which models are fast enough, and what they trade for it
For teams building phone and customer service agents. This page reads three public benchmarks with one question first: how long does the caller wait before the voice starts? Voice quality, accuracy and price are then compared among the models that are fast enough. It runs no tests of its own and combines no scores.
Short answer
- 4 models start speaking in 101 ms or less, even on slow turns (P95): Fluxions vui, Alibaba Qwen3 TTS Fast, Inworld Realtime TTS-2 Flash and Gradium TTS. On slow turns (P95), Fluxions vui is ahead on its own; Alibaba Qwen3 TTS Fast, Inworld Realtime TTS-2 Flash and Gradium TTS are a tie. On typical turns (median), Fluxions vui is ahead on its own; Gradium TTS, Inworld Realtime TTS-2 Flash and Alibaba Qwen3 TTS Fast are a tie. Gaps under 10 ms are treated as ties.
- They differ on everything else. Fluxions vui: the cheapest of the group and no public voice rating. Alibaba Qwen3 TTS Fast: not rated in the Artificial Analysis arena. Inworld Realtime TTS-2 Flash: the best on numbers and dates of the group (Speko). Gradium TTS: the most expensive of the group, the weakest on numbers and dates (Speko) and the highest Speko naturalness of the group.
- If your agent can spare about 200 ms, SpeechifyAI Simba 3.2 (median 106 ms), Inworld Realtime TTS-2 (median 153 ms), Fish Audio Drama 3 Preview (median 146 ms) and ElevenLabs Eleven v4 Turbo (median 190 ms) are rated more natural than all of them.
How much speed does a voice agent need?
People leave about 200 ms between turns in conversation. A voice agent cannot match that: before the voice starts, it has to detect that the caller stopped, transcribe, and get the first words from a language model, a step measured in hundreds of milliseconds on Speko's LLM board.
Text-to-speech is the one stage where the wait can be brought close to zero. Every 100 ms saved there can be spent on a better language model or a tool call. That is the case for ordering by speed first. It also has a limit: once the voice starts in under about 200 ms, it is rarely the slowest stage, and voice quality starts to matter more.
4 models
Headroom · P95 at or under 101 ms
Headroom to spend elsewhere in the turn.
9 models
Fast enough · P95 from 138 ms to 263 ms
Fast enough for most agents.
19 models
Visible delay · P95 of 320 ms or more
The voice becomes a visible part of the delay.
Turn gap: Stivers et al., PNAS, 2009. Tiers: Coval, 30-day window, read October 22, 2026. The breaks between tiers follow the gaps in the data: no model sits inside them.
Speed: typical turns and slow turns
The median is the typical wait. P95 is the slowest one turn in twenty, the one a caller notices. Both are shown. Differences under 10 ms are treated as ties, because the order between the leaders changes from week to week.
Headroom · P95 at or under 101 ms
Fast enough · P95 from 138 ms to 263 ms
- Alibaba Qwen3 TTS 1.7bAlibaba (Qwen) · median 104 ms 138 ms
- SpeechifyAI Simba 3.0Speechify · median 101 ms 148 ms
- SpeechifyAI Simba 3.2Speechify · median 106 ms 172 ms
- Inworld Realtime TTS-2Inworld · median 153 ms 200 ms
- Palabra TTS v1Palabra · median 109 ms 205 ms
- Fish Audio Drama 3 PreviewFish Audio · median 146 ms 213 ms
- ElevenLabs Flash v2.5ElevenLabs · median 188 ms 241 ms
- ElevenLabs Eleven v4 TurboElevenLabs · median 190 ms 244 ms
- Deepgram Flux TTSDeepgram · median 200 ms 263 ms
Visible delay · P95 of 320 ms or more
- Soniox TTS RT v1Soniox · median 260 ms 320 ms
- Soniox TTS Real-Time v2Soniox · median 263 ms 325 ms
- Rime Mist v3Rime · median 258 ms 346 ms
- Cartesia Sonic 3.6Cartesia · median 296 ms 380 ms
- Cartesia Sonic 3.5Cartesia · median 281 ms 381 ms
- Rime CodaRime · median 300 ms 415 ms
- Deepdub Phantom Z 3.4 conversationalDeepdub · median 238 ms 432 ms
- Out of Set Airy TTSOut of Set · median 216 ms 467 ms
- SpaceXAI Grok TTSxAI · median 358 ms 489 ms
- Deepgram Aura 2Deepgram · median 266 ms 526 ms
- Fish Audio S2.1 ProFish Audio · median 318 ms 568 ms
- Guava Daytona TTSGuava · median 466 ms 595 ms
- Murf AI Falcon 2Murf AI · median 286 ms 613 ms
- Smallest.ai Lightning V3.1 Pro (Jul 2026)Smallest.ai · median 356 ms 691 ms
- Alibaba Qwen3 TTS Flash RealtimeAlibaba (Qwen) · median 728 ms 789 ms
- Google Chirp 3: HDGoogle · median 534 ms 917 ms
- Deepgram Aura 2 EnDeepgram · median 520 ms 1,027 ms
- Fish Audio S2.1 Pro FreeFish Audio · median 836 ms 1,159 ms
- OpenAI GPT-4o mini TTSOpenAI · median 950 ms 5,055 ms
Time to first audio, 30-day window. Bars show P95; the median is under each name. Source: Coval, all 32 models it measures. 87 other models are not measured by Coval and have no tier.
The 4 fastest, side by side
On slow turns (P95), Fluxions vui is ahead on its own; Alibaba Qwen3 TTS Fast, Inworld Realtime TTS-2 Flash and Gradium TTS are a tie. On typical turns (median), Fluxions vui is ahead on its own; Gradium TTS, Inworld Realtime TTS-2 Flash and Alibaba Qwen3 TTS Fast are a tie. Gaps under 10 ms are treated as ties.
| Model | Median | P95 | Naturalness (Speko) | Listener Elo (Artif. Analysis) | Numbers, dates (Speko, 0 to 1) | Price per 1M chars | Multilingual (Coval tag) |
|---|---|---|---|---|---|---|---|
| Fluxions vui | 41 ms | 83 ms | not rated | not rated | not rated | $9 | Not tagged multilingual |
| Alibaba Qwen3 TTS Fastopen weights | 64 ms | 97 ms | 1536 | not rated | 0.3 | $10 | Multilingual |
| Inworld Realtime TTS-2 Flash | 61 ms | 98 ms | 1501 | 1,218 ±13 | 0.58 | $10.4 to $15 | Multilingual |
| Gradium TTS | 55 ms | 101 ms | 1592 | 1,176 ±13 | 0.27 | $35.9 to $57.8 | Multilingual |
The two voice ratings disagree. Speko's blind votes on phone-agent lines put Gradium TTS first of this group; Artificial Analysis, with a larger and more general set of listeners, puts Inworld Realtime TTS-2 Flash first. Listen to both before deciding.
Latency and tags: Coval. Naturalness and number pronunciation: Speko. Elo: Artificial Analysis. Prices: Coval, Speko, Artificial Analysis and gradium.ai/pricing (Gradium plans), as a range when sources or plans differ.
Speed against naturalness, all models
The filled points are the frontier: for each of them, no other model on Speko's board is both faster and rated more natural. Everything to the right of a frontier point and below it is a worse deal on these two measures.
| Frontier model | Time to first audio | Naturalness Elo |
|---|---|---|
| Alibaba Qwen3 TTS Fast | 37 ms | 1536 |
| Gradium TTS | 55 ms | 1592 |
| Soniox TTS Real-Time v2 | 134 ms | 1606 |
| ElevenLabs Eleven v4 Turbo | 203 ms | 1670 |
| ElevenLabs Eleven v4 | 852 ms | 1708 |
Median time to first audio, US East, and naturalness Elo for the 41 models with both, as plotted on Speko's board, read October 22, 2026. Models absent from Speko's board are not on this chart.
The same question on a second pair of sources
18 models are measured by Coval and rated in the Artificial Analysis Speech Arena. Sorted by median latency, 6 are on the frontier: no other model of the 18 is both faster and rated higher.
| Model | Median (Coval) | P95 (Coval) | Arena Elo | Arena rank | Price per 1M chars | |
|---|---|---|---|---|---|---|
| Gradium TTS | 55 ms | 101 ms | 1,176 ±13 | 23 | $35.9 to $57.8 | frontier |
| Inworld Realtime TTS-2 Flash | 61 ms | 98 ms | 1,218 ±13 | 13 | $10.4 to $15 | frontier |
| SpeechifyAI Simba 3.0 | 101 ms | 148 ms | 1,121 ±11 | 35 | $6.6 to $10 | |
| SpeechifyAI Simba 3.2 | 106 ms | 172 ms | 1,245 ±12 | 10 | $6.6 to $10 | frontier |
| Fish Audio Drama 3 Preview | 146 ms | 213 ms | 1,248 ±15 | 9 | $15 | frontier |
| Inworld Realtime TTS-2 | 153 ms | 200 ms | 1,267 ±15 | 6 | $15 to $25 | frontier |
| ElevenLabs Flash v2.5 | 188 ms | 241 ms | 1,077 ±11 | 50 | $40 to $50 | |
| ElevenLabs Eleven v4 Turbo | 190 ms | 244 ms | 1,335 ±17 | 1 | $40 | frontier |
| Deepgram Flux TTS | 200 ms | 263 ms | 1,052 ±13 | 65 | $45 | |
| Soniox TTS Real-Time v2 | 263 ms | 325 ms | 1,180 ±11 | 22 | $13 to $14.2 | |
| Cartesia Sonic 3.5 | 281 ms | 381 ms | 1,191 ±11 | 18 | $49 to $50 | |
| Murf AI Falcon 2 | 286 ms | 613 ms | 1,153 ±13 | 28 | $10 | |
| Cartesia Sonic 3.6 | 296 ms | 380 ms | 1,289 ±14 | 4 | $49 to $50 | |
| Rime Coda | 300 ms | 415 ms | 1,056 ±12 | 62 | $50 | |
| Fish Audio S2.1 Pro | 318 ms | 568 ms | 1,189 ±14 | 20 | $15 | |
| Smallest.ai Lightning V3.1 Pro (Jul 2026) | 356 ms | 691 ms | 1,169 ±12 | 25 | $19.5 | |
| SpaceXAI Grok TTS | 358 ms | 489 ms | 1,150 ±13 | 29 | $15 | |
| Google Chirp 3: HD | 534 ms | 917 ms | 1,061 ±12 | 60 | $30 |
Sources: Coval and Artificial Analysis. Elo scores carry the 95% margin shown; two scores whose intervals overlap are a tie.
Voices rated for customer service
Artificial Analysis splits its votes by use. Its Customer Service category rates 93 models; the 16 best-rated are below. Ranks whose 95% intervals overlap are statistically a tie, so read close ranks as one group.
| Rank of 93 | Model | Elo | 95% interval (±) | Median latency (Coval) |
|---|---|---|---|---|
| 1 | ElevenLabs Eleven v4 Turbo | 1379 | ±34 | 190 ms |
| 2 | ElevenLabs Eleven v4 | 1369 | ±33 | not on Coval |
| 3 | Google Gemini 3.8 Flash TTS | 1345 | ±30 | not on Coval |
| 4 | Alibaba Qwen-Audio-3.1-TTS-Plus | 1339 | ±32 | not on Coval |
| 5 | Cartesia Sonic 3.6 | 1322 | ±29 | 296 ms |
| 6 | Inworld Realtime TTS-2 | 1311 | ±30 | 153 ms |
| 7 | Google Gemini 3.8 Flash-Lite TTS | 1308 | ±28 | not on Coval |
| 8 | Fish Audio Drama 3 Preview | 1301 | ±31 | 146 ms |
| 9 | Alibaba Qwen-Audio-3.0-TTS-Plus | 1284 | ±28 | not on Coval |
| 10 | BreezeBlue Breeze TTS 2 | 1273 | ±28 | not on Coval |
| 11 | ElevenLabs Eleven v3 Conversational | 1268 | ±26 | not on Coval |
| 12 | Mistral Voxtral 3 TTS | 1264 | ±31 | not on Coval |
| 13 | VUI Labs Luna TTS | 1254 | ±25 | not on Coval |
| 14 | StepFun StepAudio 2.5 TTS (Aug 2026) | 1249 | ±29 | not on Coval |
| 15 | SpeechifyAI Simba 3.2 | 1248 | ±25 | 106 ms |
| 16 | Gradium TTS | 1242 | ±27 | 55 ms |
Source: Artificial Analysis, Customer Service category, read October 22, 2026. Latency from Coval where the model is measured.
Which model for which need
| If the priority is | Look first at | Because | Check before choosing |
|---|---|---|---|
| Lowest latency in several languages, voice quality first | Gradium TTS | Median 55 ms, P95 101 ms; Speko naturalness 1592, the highest of the multilingual models in the fastest tier | the most expensive of the group and the weakest on numbers and dates (Speko) |
| Lowest latency in several languages, price first | Alibaba Qwen3 TTS Fast | Median 64 ms; list price $10 per 1M characters | not rated in the Artificial Analysis arena |
| Lowest latency at the lowest price | Fluxions vui | Median 41 ms, $9 per 1M characters | no public voice rating |
| Low latency with open weights | Alibaba Qwen3 TTS Fast | P95 97 ms, open-weights licence, $10 per 1M characters | not rated in the Artificial Analysis arena |
| The best-rated voice, if about 200 ms is acceptable | ElevenLabs Eleven v4 Turbo | Arena Elo 1,335 ±17; median 190 ms | — |
| Amounts, dates and reference numbers read aloud | ElevenLabs Eleven v4 | Speko pronunciation score 0.9, tied highest of 30 with ElevenLabs Eleven v4 Turbo | — |
Why accuracy is not ranked here
Coval measures word error rate by transcribing the generated speech. On its 30-day table, 25 of 32 models fall between 1.8% and 2.8%. The field is too close to rank on this measure. Speko's number and date test separates models far more, so that is the accuracy measure used on this site.
Check it yourself
- Coval's TTS dashboard: set the window to 30 days and switch the percentile between P50 and P95.
- Speko's TTS board: set the axes to time to first audio and naturalness, and turn on the Pareto frontier.
- The Artificial Analysis leaderboard: set Category to Customer Service.
Questions and answers
- What is the fastest TTS API for voice agents?
- On Coval's 30-day window read on October 22, 2026, 4 models start speaking in 101 ms or less even on slow turns (P95): Fluxions vui (83 ms), Alibaba Qwen3 TTS Fast (97 ms), Inworld Realtime TTS-2 Flash (98 ms) and Gradium TTS (101 ms). Fluxions vui is ahead on its own by more than 10 ms. On the median, Fluxions vui (41 ms) leads.
- What is the fastest multilingual TTS API?
- Among the models Coval tags multilingual, Gradium TTS (median 55 ms), Inworld Realtime TTS-2 Flash (median 61 ms) and Alibaba Qwen3 TTS Fast (median 64 ms) are a tie on the median, within 10 ms.
- Which fast TTS sounds the most natural?
- Among the fastest tier, Gradium TTS has the highest Speko naturalness (1592). Artificial Analysis rates Inworld Realtime TTS-2 Flash highest of the same group (1,218 ±13). If the agent can spare about 200 ms, SpeechifyAI Simba 3.2, Inworld Realtime TTS-2, Fish Audio Drama 3 Preview and ElevenLabs Eleven v4 Turbo are rated more natural than all of them.
- Which low-latency TTS reads numbers and dates best?
- In the fastest tier, Inworld Realtime TTS-2 Flash scores 0.58 on Speko's number and date test. Across all 30 models Speko scores, the top score is 0.9, for ElevenLabs Eleven v4 and ElevenLabs Eleven v4 Turbo.
- What does low-latency TTS cost?
- List prices per 1M characters in the fastest tier: Fluxions vui $9, Alibaba Qwen3 TTS Fast $10, Inworld Realtime TTS-2 Flash $10.4 to $15 and Gradium TTS $35.9 to $57.8. Prices are list prices; plans differ.
- Do these figures hold outside the United States?
- Not necessarily. All latency here is measured from US East, in English; no public source measures from Europe.
Method
- This page runs no tests. Every number is copied from one of three public benchmarks, read on the date shown, with a link.
- It is written for real-time agents, so speed is the first filter. A page about narration would order these models differently.
- Ties: latencies within 10 ms, and Elo scores whose published intervals overlap, are treated as ties.
- No score thresholds are used to pick which models appear. Each comparison shows every model the sources share.
- Ranks from different benchmarks are never added together. Full methodology.
Limits
- The benchmarks update daily; the order among the fastest models changes within a week.
- Coverage differs: Coval measures 32 models, Speko's board plots 41, Artificial Analysis rates 98.
- Prices are list prices, and plans differ.
- English, US East. No public measurement from Europe exists.
Data read October 22, 2026. Published by Rank-ly. Corrections: contact. Data from Coval, Speko and Artificial Analysis belongs to its publishers.