Best TTS for voice agents
| # | Model | Scoreof 100 | Covalfirst audio, p50 | Spekop50 | Humanness Indexmedian | Price$ / 1M chars |
|---|---|---|---|---|---|---|
| 1 | Simba 3.2SpeechifyAI | 87 | 101 ms | 85 ms | 428 ms | $10.00 |
| 2 | Sonic 3.6Cartesia | 83 | 288 ms | 120 ms | 101 ms | $49.00 |
| 3 | Eleven v4 TurboElevenLabs | 70 | 191 ms | 205 ms | 158 ms | $40.00 |
| 4 | Realtime TTS-2Inworld | 70 | 161 ms | 116 ms | 288 ms | $20.83 |
| 5 | Simba 3.5SpeechifyAI | 67 | 90 ms | — | — | $15.00 |
| 6 | Realtime TTS-2 FlashInworld | 58 | 59 ms | 82 ms | — | $10.42 |
| 7 | AssemblyAI-TTS-1AssemblyAI | 58 | 204 ms | — | — | $15.00 |
| 8 | Falcon 2Murf AI | 57 | 141 ms | — | — | $10.00 |
| 9 | Soniox TTS Real-Time v2Soniox | 53 | 304 ms | 381 ms | — | $14.23 |
| 10 | Simba 3.0SpeechifyAI | 42 | 105 ms | — | — | $10.00 |
| 11 | Flash v2.5ElevenLabs | 36 | 200 ms | — | 197 ms | $50.00 |
| 12 | Chirp 3: HDGoogle | 26 | 514 ms | — | — | $30.00 |
Quality first, then speed. Latency is time to first audio from each source, never averaged. "—" means no figure from that source. Sorted by speed instead.
Best score at each speed limit
| If you need | Best-scoring model | Scoreof 100 | First audioCoval p50 |
|---|---|---|---|
| Up to 100 ms | Simba 3.5SpeechifyAI | 67 | 90 ms |
| Up to 150, 200 or 300 ms | Simba 3.2SpeechifyAI | 87 | 101 ms |
Simba 3.5 on the other sources: Speko no figure, Humanness Index no figure.
4 models score 70 or more and have a Coval figure: Simba 3.2 (101 ms), Realtime TTS-2 (161 ms), Eleven v4 Turbo (191 ms), Sonic 3.6 (288 ms).
What this does not cover
Across the 17 models, Coval medians run from 59 ms to 719 ms, with the middle one at 200 ms. Time to first audio is only the TTS share of the delay in a voice agent: speech recognition and the language model add their own. We do not test interruption handling, streaming support or phone-line audio quality either, so use this to build a shortlist and test the shortlist.
Questions people ask
- What is the best TTS for voice agents?
- Simba 3.2 (SpeechifyAI) has the highest score among the 17 models Coval measures: 87, with a median first audio of 101 ms. If you need first audio within 100 ms, the best score is 67 (Simba 3.5). 4 models score 70 or more and have a Coval figure: Simba 3.2 (101 ms), Realtime TTS-2 (161 ms), Eleven v4 Turbo (191 ms), Sonic 3.6 (288 ms).
- What time to first audio can I expect from a TTS API?
- Across the 17 models, Coval medians run from 59 ms to 719 ms, with the middle one at 200 ms. Other sources measure differently, so read their numbers separately.
- Does TTS latency alone decide how fast a voice agent feels?
- No. The wait you hear also includes speech recognition and the language model. Time to first audio is the text-to-speech share of it, and it is the only part this page shows.
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.