Fastest text to speech API
Coval, our main source
| # | Model | First audioCoval p50 | Scoreof 100 | Price$ / 1M chars |
|---|---|---|---|---|
| 1 | Realtime TTS-2 FlashInworld | 59 ms | 58 | $10.42 |
| 2 | Simba 3.5SpeechifyAI | 90 ms | 67 | $15.00 |
| 3 | CodaRime | 96 ms | 23 | $50.00 |
| 4 | Simba 3.2SpeechifyAI | 101 ms | 87 | $10.00 |
| 5 | Simba 3.0SpeechifyAI | 105 ms | 42 | $10.00 |
| 6 | Falcon 2Murf AI | 141 ms | 57 | $10.00 |
| 7 | Realtime TTS-2Inworld | 161 ms | 70 | $20.83 |
| 8 | Eleven v4 TurboElevenLabs | 191 ms | 70 | $40.00 |
| 9 | Flash v2.5ElevenLabs | 200 ms | 36 | $50.00 |
| 10 | AssemblyAI-TTS-1AssemblyAI | 204 ms | 58 | $15.00 |
Coval: Open-source harness run every 30 minutes from us-east-1; median over a rolling 7-day window. Streaming; connection handshakes excluded; counts leading silence until the first audible sample. Showing the fastest 10.
Speko
| # | Model | First audioSpeko p50 | Scoreof 100 | Price$ / 1M chars |
|---|---|---|---|---|
| 1 | Realtime TTS-2 FlashInworld | 82 ms | 58 | $10.42 |
| 2 | Simba 3.2SpeechifyAI | 85 ms | 87 | $10.00 |
| 3 | Gradium TTSGradium | 105 ms | 10 | $47.20 |
| 4 | Realtime TTS-2Inworld | 116 ms | 70 | $20.83 |
| 5 | Sonic 3.6Cartesia | 120 ms | 83 | $49.00 |
| 6 | Sonic 3.5Cartesia | 121 ms | 73 | $49.00 |
| 7 | Lightning v3.1Smallest.ai | 173 ms | 17 | $25.00 |
| 8 | Fish Audio S2.1 ProFish Audio | 185 ms | 15 | $15.00 |
| 9 | Eleven v4 TurboElevenLabs | 205 ms | 70 | $40.00 |
| 10 | v3 ConversationalElevenLabs | 266 ms | 71 | $50.00 |
Speko: Sequential requests from US East, about 30 per run, re-run per model on the date shown. Showing the fastest 10.
Humanness Index
| # | Model | First audioHumanness Index median | Scoreof 100 | Price$ / 1M chars |
|---|---|---|---|---|
| 1 | Sonic 3.6Cartesia | 101 ms | 83 | $49.00 |
| 2 | SonicCartesia | 116 ms | 0 | — |
| 3 | Falcon 2.1Murf AI | 117 ms | 48 | $10.00 |
| 4 | Sonic 3.5Cartesia | 128 ms | 73 | $49.00 |
| 5 | Fish Audio S2.1 ProFish Audio | 141 ms | 15 | $15.00 |
| 6 | Eleven v4 TurboElevenLabs | 158 ms | 70 | $40.00 |
| 7 | Sonic 2Cartesia | 159 ms | 3 | — |
| 8 | Sonic 3Cartesia | 166 ms | 31 | $49.00 |
| 9 | Flash v2.5ElevenLabs | 197 ms | 36 | $50.00 |
| 10 | Neuphonic TTSNeuphonic | 276 ms | 9 | $20.80 |
Humanness Index: Region not stated. 50 sequential streaming API requests per model, measured once (the date shown). Showing the fastest 10.
Fast is not the same as good
Realtime TTS-2 Flash (Inworld) is quickest on Coval, at 59 ms, but it scores 58. The fastest model that scores 70 or more is Simba 3.2 at 101 ms (score 87).
Why the sources disagree
Coval puts Realtime TTS-2 Flash first (59 ms); Speko puts Realtime TTS-2 Flash first (82 ms); Humanness Index puts Sonic 3.6 first (101 ms).
Simba 3.2 shows the problem: Speko has it at 85 ms and Humanness Index at 428 ms, 5× apart for the same model.
The sources use different harnesses and regions, so we show them side by side and never average them. Latency is not part of the 0–100 score. Of 121 models on the leaderboard, 39 have a figure from at least one source.
Questions people ask
- What is the fastest text-to-speech API?
- Coval puts Realtime TTS-2 Flash first (59 ms); Speko puts Realtime TTS-2 Flash first (82 ms); Humanness Index puts Sonic 3.6 first (101 ms). Each source runs its own test, so read the table for the source you trust and do not average them.
- Is the fastest TTS API also a good one?
- Realtime TTS-2 Flash (Inworld) is quickest on Coval, at 59 ms, but it scores 58. The fastest model that scores 70 or more is Simba 3.2 at 101 ms (score 87).
- Why do latency numbers differ between benchmarks?
- Harness, region and sample size differ. Coval: Open-source harness run every 30 minutes from us-east-1; median over a rolling 7-day window. Streaming; connection handshakes excluded; counts leading silence until the first audible sample. Speko: Sequential requests from US East, about 30 per run, re-run per model on the date shown. Humanness Index: Region not stated. 50 sequential streaming API requests per model, measured once (the date shown). Simba 3.2 shows the problem: Speko has it at 85 ms and Humanness Index at 428 ms, 5× apart for the same model.
Where the numbers come from
Quality is the 0–100 score from our rank aggregation of 3 public boards: artificialanalysis.ai (Speech Arena Elo), voicearena.com (arena Elo, US English) and humannessindex.vapi.ai (humanness score 0–100). Prices are dollars per 1M characters, taken from Artificial Analysis and provider pages; per-minute prices are converted to the same unit. Latency is time to first audio (the wait before sound starts), copied from Coval, Speko and Humanness Index and never averaged across them. We do not measure anything ourselves. How latency is sourced.