Fish Audio · read on October 27, 2026
Fish Audio Drama 3 Preview
Fish Audio Drama 3 Preview is a TTS (text-to-speech) model from Fish Audio. Coval measures its P95 time to first audio at 209.3 ms, which puts it in the Fast enough speed tier. It ranks 10 of 86 in the full ranking.
- Speed tier
- Fast enough
- P95 / median
- 209.3 ms / 146 ms
- Listener Elo
- 1,245
- Price per 1M chars
- $15
Where it ranks
Full ranking: 10 of 86. Use case ranks, by speed:
- Customer service: 6 of 70
- Voice agents: 10 of 13
- Assistants: 6 of 69
- Entertainment: 6 of 69
- Knowledge sharing: 6 of 73
The tier and the ranks are computed by this site from the benchmarks' figures, not declared by Fish Audio.
Measurements and sources
Coval
| Median time to first audio (P50) | 146 ms |
| Time to first audio, P95 | 209.3 ms |
| Time to first audio, mean | 165.6 ms |
| Time to first audio, P90 | 183.2 ms |
| Timed requests | 1,402 |
Artificial Analysis
| Listener Elo | 1,245 |
| Listener Elo, 95% interval | 14 |
| Listener Elo, customer service | 1,296 |
| Hard passages pronounced correctly | 93.2% |
Cekura
| Time to first audio, streamed text | 312 ms |
| Time to first audio, text sent whole | 321 ms |
| Reading accuracy | 91.3% |
| Estimated naturalness (1 to 5) | 4.54 |
Details
| Provider | Fish Audio |
| Access | Open weights |
| Streaming | A benchmark measures its time to first audio |
| Languages tested | German, Chinese |
| Measured by | Artificial Analysis, Cekura, Coval |
| Official site | fish.audio |
Not collected
Nothing below is estimated. Each line says which measure is missing and why.
- Naturalness (Speko naturalness): not measured by Speko (reading of 2026-10-27)
- Numbers and dates (Speko pronunciation score): not measured by Speko (reading of 2026-10-27)
Sources
- https://artificialanalysis.ai/text-to-speech (read on October 27, 2026)
- https://artificialanalysis.ai/text-to-speech/leaderboard/provider-voice (read on October 27, 2026)
- https://benchmarks.cekura.ai/text-to-speech (read on October 27, 2026)
- https://benchmarks-api-6wxgp27p2a-ue.a.run.app/v1/results/aggregates?benchmark=TTS&window=30d&include_series=false (read on October 27, 2026)
Values are copied from these pages on the date shown. The match between this model and its benchmark entries is inferred, because a source writes the name differently: Cekura: “Fish Audio Drama 3”; Coval: “Fish Audio drama3-preview”. See the methodology.
Compare it with the reference model: Gradium TTS vs Fish Audio Drama 3 Preview.