Data read on November 19, 2026
Gradium TTS latency
Time to first audio is the delay between sending text and hearing the first sound. For a voice agent, it is the part of the turn the caller notices.
| Latency | Gradium TTS | Rank (median) |
|---|---|---|
| Median time to first audio | 64 ms | 5 / 32 |
| P90 | 161 ms | |
| P95 time to first audio (slow turns) | 172 ms | 11 / 32 |
| Leading silence (median) | 0 ms | 2 / 32 |
| Word error rate (mean) | 1.7% | 16 / 32 |
| Samples | 96 |
Source : Coval · Read on November 19, 2026 · gradium-tts-beta-202609
Coval TTS leaderboard, ordered by median time to first audio
Every model Coval measures, as published on its public board. Gradium is highlighted.
| # | Model | Median | P95 | Leading silence | WER |
|---|---|---|---|---|---|
| 1 | Fluxions vui | 36 ms | 65 ms | 8 ms | 2.1% |
| 2 | Deepdub dd-etts-3.3 | 56 ms | 86 ms | 2 ms | 1.6% |
| 3 | Inworld inworld-tts-2-flash | 60 ms | 118 ms | 0 ms | 1.7% |
| 4 | Nari Labs qwen3-tts-fast | 63 ms | 101 ms | 10 ms | 1.9% |
| 5 | Gradium gradium-tts-beta-202609 | 64 ms | 172 ms | 0 ms | 1.7% |
| 6 | Palabra palabra-tts-v1 | 65 ms | 135 ms | 0 ms | 1.7% |
| 7 | Speechify simba-3.5 | 93 ms | 143 ms | 11 ms | 2.7% |
| 8 | Rime coda | 99 ms | 153 ms | 24 ms | 1.8% |
| 9 | Speechify simba-3.0 | 102 ms | 151 ms | 10 ms | 2.2% |
| 10 | Speechify simba-3.2 | 103 ms | 149 ms | 10 ms | 2.0% |
| 11 | Baseten qwen3-tts-1.7b | 118 ms | 143 ms | 37 ms | 1.8% |
| 12 | Murf falcon-2 | 141 ms | 235 ms | 31 ms | 1.6% |
| 13 | Inworld inworld-tts-2 | 160 ms | 213 ms | 57 ms | 1.5% |
| 14 | ElevenLabs eleven_flash_v2_5 | 201 ms | 546 ms | 41 ms | 1.9% |
| 15 | ElevenLabs eleven_v4_turbo | 203 ms | 4291 ms | 64 ms | 1.7% |
| 16 | Deepgram flux-haley-en | 205 ms | 449 ms | 53 ms | 1.9% |
| 17 | Assemblyai assemblyai-tts-1 | 209 ms | 328 ms | 121 ms | 1.9% |
| 18 | Airy airy-tts-v1 | 216 ms | 585 ms | 5 ms | 1.6% |
| 19 | Guava daytona-tts | 261 ms | 297 ms | 3 ms | 1.5% |
| 20 | Xai grok-tts | 262 ms | 454 ms | 50 ms | 1.7% |
| 21 | Deepgram aura-2-thalia-en | 262 ms | 570 ms | 126 ms | 1.9% |
| 22 | Smallest AI lightning_v3.1_pro | 284 ms | 531 ms | 56 ms | 1.6% |
| 23 | Cartesia sonic-3.6 | 292 ms | 398 ms | 138 ms | 1.6% |
| 24 | Soniox tts-rt-v2 | 301 ms | 353 ms | 81 ms | 3.0% |
| 25 | Soniox tts-rt-v1 | 305 ms | 363 ms | 75 ms | 2.9% |
| 26 | Fish Audio s2.1-pro | 354 ms | 502 ms | 4 ms | 1.6% |
| 27 | Fish Audio drama-3-preview-us-east | 367 ms | 485 ms | 71 ms | 1.4% |
| 28 | Rime mistv3 | 402 ms | 517 ms | 135 ms | 2.3% |
| 29 | Cloudflare aura-2-en | 496 ms | 684 ms | 250 ms | 2.1% |
| 30 | Google chirp-3-hd | 532 ms | 883 ms | 394 ms | 1.7% |
| 31 | Alibaba qwen3-tts-flash-realtime | 719 ms | 763 ms | 91 ms | 2.7% |
| 32 | Fish Audio s2.1-pro-free | 819 ms | 1100 ms | 5 ms | 1.6% |
Official Gradium page
Coval leaderboard reading cited on gradium.ai/api and llms.txt
47 ms · P50 time to first audio (7-day window, read September 24, 2026)
read September 24, 2026
Text-to-Speech product page figures
49 ms · TTFA P50 (Speko TTS benchmark)
50 ms · TTFA P50 (Perceived TTFA, Coval benchmark)
5 ms · Latency spread (P75 − P25) (tightest of the 5 models tested, over 480 runs)
81.0% · Hard-case pass rate (500 sentences, 5 languages)
read 2026-11-19