28 paid TTS APIs, each with its documented default voice, on the same prompts as the local board. Kept off the local board by default; the Scores page can show them side by side (default voice only). First speech = ms until audible speech on a warm connection, measured from Eastern US (2026-11-01 → 2026-11-04); hover it for p90 and call counts. UTMOS, WER, SIM and Health mean the same as on the Scores page: the mean over every clip (3 takes of each prompt). Prices are list $/1M characters; * = hover for a note.
Every take of every prompt, nothing cherry-picked.
Prompt 1 [en] "Open the browser and read my email."
Model
Take 1
Take 2
Take 3
ElevenLabs Eleven v4
ElevenLabs Eleven v4 Turbo
Rime Coda
Giggy TTS
Cartesia Sonic 3.6
Gemini 3.8 Flash TTS
Inworld TTS-2
Speechify SIMBA 3.5
Fish Audio Drama 3
Mistral Voxtral TTS 3
xAI TTS
AssemblyAI TTS-1
Soniox TTS v2
StepFun Step-Audio 3 TTS
Smallest.ai Lightning v3.1 Pro
Deepgram Flux TTS
Microsoft MAI-Voice-2.1
VUI Labs Luna
OpenAI gpt-4o-mini-tts
Murf Falcon 2.1
HeyGen Voice 1
MiniMax speech-2.8-hd
OpenAI gpt-realtime-2.1
Bland Speech v3
Gradium
Async Flash v1.5
Palabra TTS
Resemble Ultra
Prompt 2 [en] "I'll start a new git branch, push the changes, and open a pull request when the tests pass…"
Model
Take 1
Take 2
Take 3
ElevenLabs Eleven v4
ElevenLabs Eleven v4 Turbo
Rime Coda
Giggy TTS
Cartesia Sonic 3.6
Gemini 3.8 Flash TTS
Inworld TTS-2
Speechify SIMBA 3.5
Fish Audio Drama 3
Mistral Voxtral TTS 3
xAI TTS
AssemblyAI TTS-1
Soniox TTS v2
StepFun Step-Audio 3 TTS
Smallest.ai Lightning v3.1 Pro
Deepgram Flux TTS
Microsoft MAI-Voice-2.1
VUI Labs Luna
OpenAI gpt-4o-mini-tts
Murf Falcon 2.1
HeyGen Voice 1
MiniMax speech-2.8-hd
OpenAI gpt-realtime-2.1
Bland Speech v3
Gradium
Async Flash v1.5
Palabra TTS
Resemble Ultra
Prompt 3 [en] "The Parakeet TDT zero point six billion parameter model achieves one point six nine percen…"
Model
Take 1
Take 2
Take 3
ElevenLabs Eleven v4
ElevenLabs Eleven v4 Turbo
Rime Coda
Giggy TTS
Cartesia Sonic 3.6
Gemini 3.8 Flash TTS
Inworld TTS-2
Speechify SIMBA 3.5
Fish Audio Drama 3
Mistral Voxtral TTS 3
xAI TTS
AssemblyAI TTS-1
Soniox TTS v2
StepFun Step-Audio 3 TTS
Smallest.ai Lightning v3.1 Pro
Deepgram Flux TTS
Microsoft MAI-Voice-2.1
VUI Labs Luna
OpenAI gpt-4o-mini-tts
Murf Falcon 2.1
HeyGen Voice 1
MiniMax speech-2.8-hd
OpenAI gpt-realtime-2.1
Bland Speech v3
Gradium
Async Flash v1.5
Palabra TTS
Resemble Ultra
Prompt 4 [en] "Run pytest tests slash test underscore voice dot py with verbose flag and capture flag set…"
Model
Take 1
Take 2
Take 3
ElevenLabs Eleven v4
ElevenLabs Eleven v4 Turbo
Rime Coda
Giggy TTS
Cartesia Sonic 3.6
Gemini 3.8 Flash TTS
Inworld TTS-2
Speechify SIMBA 3.5
Fish Audio Drama 3
Mistral Voxtral TTS 3
xAI TTS
AssemblyAI TTS-1
Soniox TTS v2
StepFun Step-Audio 3 TTS
Smallest.ai Lightning v3.1 Pro
Deepgram Flux TTS
Microsoft MAI-Voice-2.1
VUI Labs Luna
OpenAI gpt-4o-mini-tts
Murf Falcon 2.1
HeyGen Voice 1
MiniMax speech-2.8-hd
OpenAI gpt-realtime-2.1
Bland Speech v3
Gradium
Async Flash v1.5
Palabra TTS
Resemble Ultra
Prompt 5 [fr] "Bonjour, je m'appelle Cicero et je vais vous aider avec votre code aujourd'hui."
Model
Take 1
Take 2
Take 3
ElevenLabs Eleven v4
ElevenLabs Eleven v4 Turbo
Rime Coda
Giggy TTS
not offered in fr
Cartesia Sonic 3.6
Gemini 3.8 Flash TTS
Inworld TTS-2
Speechify SIMBA 3.5
Fish Audio Drama 3
Mistral Voxtral TTS 3
xAI TTS
AssemblyAI TTS-1
Soniox TTS v2
StepFun Step-Audio 3 TTS
Smallest.ai Lightning v3.1 Pro
Deepgram Flux TTS
Microsoft MAI-Voice-2.1
VUI Labs Luna
not offered in fr
OpenAI gpt-4o-mini-tts
Murf Falcon 2.1
HeyGen Voice 1
MiniMax speech-2.8-hd
not offered in fr
OpenAI gpt-realtime-2.1
Bland Speech v3
not offered in fr
Gradium
Async Flash v1.5
Palabra TTS
Resemble Ultra
Prompt 6 [es] "¿Este modelo es bilingüe? Añade el español, corrige la pronunciación automática y reinicia…"
Model
Take 1
Take 2
Take 3
ElevenLabs Eleven v4
ElevenLabs Eleven v4 Turbo
Rime Coda
Giggy TTS
not offered in es
Cartesia Sonic 3.6
Gemini 3.8 Flash TTS
Inworld TTS-2
Speechify SIMBA 3.5
Fish Audio Drama 3
Mistral Voxtral TTS 3
xAI TTS
AssemblyAI TTS-1
Soniox TTS v2
StepFun Step-Audio 3 TTS
Smallest.ai Lightning v3.1 Pro
Deepgram Flux TTS
Microsoft MAI-Voice-2.1
VUI Labs Luna
not offered in es
OpenAI gpt-4o-mini-tts
Murf Falcon 2.1
HeyGen Voice 1
MiniMax speech-2.8-hd
not offered in es
OpenAI gpt-realtime-2.1
Bland Speech v3
Gradium
Async Flash v1.5
Palabra TTS
Resemble Ultra
Cloning · fidelity, naturalness, intelligibility
Each product clones the same consented voice through its own API. Sample lengths differ on purpose: each gets the longest cut it accepts, because some need more audio for a good clone. SIM = speaker similarity to the reference (WavLM, same as the local board), higher is closer.
Reference voicefull recording, consented by the speaker
Top 15 · metric:localhosted API
Model
SIM ↑
UTMOS ↑
WER ↓
Health
Sample sent
Vendor min
Recommended
Clone langs
Inworld TTS-2inworld-tts-2
0.598
4.337
0.062
✓
24.8 s
3 s
10–30 s
en es fr
Fish Audio Drama 3drama-3
0.632
4.299
0.035
✓
24.8 s
—
10–30 s
en es fr
StepFun Step-Audio 3 TTSstepaudio-3-tts
0.686
4.412
0.074
✓
9.8 s
5 s
5–10 s
en
Smallest.ai Lightning v3.1 Prolightning_v3.1_pro
0.754
4.073
0.042
✓
14.1 s
5 s
5–15 s
en es fr
VUI Labs Lunaluna-tts
0.853
4.342
0.073
✓
58.3 s
3 s
3–60 s
en
Bland Speech v3btts-3
0.821
4.054
0.091
✓
58.3 s
—
≈10 s
en es
Gradiumdefault · vendor picks
0.741
4.438
0.093
✓
112.5 s
10 s
≥10 s
en es fr
Async Flash v1.5async_flash_v1.5
0.706
4.036
0.042
✓
112.5 s
3 s
—
en es fr
Soniox TTS v2tts-rt-v2
0.620
4.144
0.043
✓
112.5 s
—
≤120 s
en es fr
Mistral Voxtral TTS 3voxtral-mini-tts-3-0
0.737
4.262
0.049
✓
58.3 s
a few s
—
en es fr
HeyGen Voice 1heygen-voice-1
0.625
4.297
0.061
✓
112.5 s
—
60–180 s
en es fr
ElevenLabs Eleven v4eleven_v4
0.765
4.006
0.045
✓
112.5 s
10 s
60–120 s
en es fr
ElevenLabs Eleven v4 Turboeleven_v4_turbo
0.746
3.967
0.049
✓
112.5 s
10 s
60–120 s
en es fr
xAI TTSxai-tts
0.578
4.108
0.038
✓
112.5 s
—
≥90 s
en es fr
Cartesia Sonic 3.6sonic-3.6
0.769
4.201
0.051
✓
58.3 s
10 s
≤60 s
en es fr
Speechify SIMBA 3.5simba-3.5
0.801
4.237
0.038
✓
24.8 s
10 s
10–30 s
en es fr
Gemini 3.8 Flash TTSgemini-3.8-flash-tts
0.658
3.872
0.030
✓
24.8 s
10 s
10–30 s
en es fr
Palabra TTSauto · vendor picks
0.656
3.842
0.081
✓
112.5 s
30 s
—
en es fr
Resemble Ultraresemble-ultra
0.563
4.175
0.045
✓
58.3 s
5 s
≤60 s
en es fr
Giggy TTSgiggyspeech
0.792
4.216
0.087
✓
9.8 s
3 s
3–10 s
en
AssemblyAI TTS-1assemblyai-tts-1
0.646
4.417
0.081
✓
19 s
3 s
10–20 s
en
Not in the cloning lens (7: no API cloning on our route)
Rime Coda: enterprise-trained voices only (30–60 min of audio, days)
Deepgram Flux TTS: no voice cloning
Microsoft MAI-Voice-2.1: needs Microsoft Limited Access approval
OpenAI gpt-4o-mini-tts: custom voices are limited to approved customers
Murf Falcon 2.1: API cloning is Enterprise-plan only
MiniMax speech-2.8-hd: no cloning through OpenRouter (our route)
OpenAI gpt-realtime-2.1: custom voices are limited to approved customers
Listen: cloned voice
Prompt 1 [en] "Open the browser and read my email."
Model
Take 1
Take 2
Take 3
ElevenLabs Eleven v4
ElevenLabs Eleven v4 Turbo
Giggy TTS
Cartesia Sonic 3.6
Gemini 3.8 Flash TTS
Inworld TTS-2
Speechify SIMBA 3.5
Fish Audio Drama 3
Mistral Voxtral TTS 3
xAI TTS
AssemblyAI TTS-1
Soniox TTS v2
StepFun Step-Audio 3 TTS
Smallest.ai Lightning v3.1 Pro
VUI Labs Luna
HeyGen Voice 1
Bland Speech v3
Gradium
Async Flash v1.5
Palabra TTS
Resemble Ultra
Prompt 2 [en] "I'll start a new git branch, push the changes, and open a pull request when the tests pass…"
Model
Take 1
Take 2
Take 3
ElevenLabs Eleven v4
ElevenLabs Eleven v4 Turbo
Giggy TTS
Cartesia Sonic 3.6
Gemini 3.8 Flash TTS
Inworld TTS-2
Speechify SIMBA 3.5
Fish Audio Drama 3
Mistral Voxtral TTS 3
xAI TTS
AssemblyAI TTS-1
Soniox TTS v2
StepFun Step-Audio 3 TTS
Smallest.ai Lightning v3.1 Pro
VUI Labs Luna
HeyGen Voice 1
Bland Speech v3
Gradium
Async Flash v1.5
Palabra TTS
Resemble Ultra
Prompt 3 [en] "The Parakeet TDT zero point six billion parameter model achieves one point six nine percen…"
Model
Take 1
Take 2
Take 3
ElevenLabs Eleven v4
ElevenLabs Eleven v4 Turbo
Giggy TTS
Cartesia Sonic 3.6
Gemini 3.8 Flash TTS
Inworld TTS-2
Speechify SIMBA 3.5
Fish Audio Drama 3
Mistral Voxtral TTS 3
xAI TTS
AssemblyAI TTS-1
Soniox TTS v2
StepFun Step-Audio 3 TTS
Smallest.ai Lightning v3.1 Pro
VUI Labs Luna
HeyGen Voice 1
Bland Speech v3
Gradium
Async Flash v1.5
Palabra TTS
Resemble Ultra
Prompt 4 [en] "Run pytest tests slash test underscore voice dot py with verbose flag and capture flag set…"
Model
Take 1
Take 2
Take 3
ElevenLabs Eleven v4
ElevenLabs Eleven v4 Turbo
Giggy TTS
Cartesia Sonic 3.6
Gemini 3.8 Flash TTS
Inworld TTS-2
Speechify SIMBA 3.5
Fish Audio Drama 3
Mistral Voxtral TTS 3
xAI TTS
AssemblyAI TTS-1
Soniox TTS v2
StepFun Step-Audio 3 TTS
Smallest.ai Lightning v3.1 Pro
VUI Labs Luna
HeyGen Voice 1
Bland Speech v3
Gradium
Async Flash v1.5
Palabra TTS
Resemble Ultra
Prompt 5 [fr] "Bonjour, je m'appelle Cicero et je vais vous aider avec votre code aujourd'hui."
Model
Take 1
Take 2
Take 3
ElevenLabs Eleven v4
ElevenLabs Eleven v4 Turbo
Giggy TTS
not offered in fr
Cartesia Sonic 3.6
Gemini 3.8 Flash TTS
Inworld TTS-2
Speechify SIMBA 3.5
Fish Audio Drama 3
Mistral Voxtral TTS 3
xAI TTS
AssemblyAI TTS-1
not offered in fr
Soniox TTS v2
StepFun Step-Audio 3 TTS
not offered in fr
Smallest.ai Lightning v3.1 Pro
VUI Labs Luna
not offered in fr
HeyGen Voice 1
Bland Speech v3
not offered in fr
Gradium
Async Flash v1.5
Palabra TTS
Resemble Ultra
Prompt 6 [es] "¿Este modelo es bilingüe? Añade el español, corrige la pronunciación automática y reinicia…"
Model
Take 1
Take 2
Take 3
ElevenLabs Eleven v4
ElevenLabs Eleven v4 Turbo
Giggy TTS
not offered in es
Cartesia Sonic 3.6
Gemini 3.8 Flash TTS
Inworld TTS-2
Speechify SIMBA 3.5
Fish Audio Drama 3
Mistral Voxtral TTS 3
xAI TTS
AssemblyAI TTS-1
not offered in es
Soniox TTS v2
StepFun Step-Audio 3 TTS
not offered in es
Smallest.ai Lightning v3.1 Pro
VUI Labs Luna
not offered in es
HeyGen Voice 1
Bland Speech v3
Gradium
Async Flash v1.5
Palabra TTS
Resemble Ultra
Warm calls on one fixed English sentence, fastest first. Delivery p3 = the full prompt-3 clip including vendor pacing (some stream at about playback speed), so don't rank by it. † = sweep calls only.
Model
Transport
Median ms
p90 ms
Calls · errors
Delivery p3
Sample rate
AssemblyAI TTS-1
transport websocket
median 30
p90 51
calls · errors 36 · 0
delivery 1.5 s
SR 24 kHz
Gradium
transport http
median 53
p90 63
calls · errors 36 · 0
delivery 5.7 s
SR 24 kHz
Palabra TTS
transport websocket
median 77
p90 81
calls · errors 27 · 0
delivery 1.1 s
SR 48 kHz
Speechify SIMBA 3.5
transport http
median 88
p90 106
calls · errors 36 · 0
delivery 0.8 s
SR 24 kHz
Inworld TTS-2
transport http
median 98
p90 121
calls · errors 39 · 0
delivery 3.1 s
SR 48 kHz
Rime Coda
transport http
median 108
p90 179
calls · errors 36 · 0
delivery 4.6 s
SR 24 kHz
Cartesia Sonic 3.6
transport websocket
median 125
p90 184
calls · errors 36 · 0
delivery 2.1 s
SR 24 kHz
Murf Falcon 2.1
transport http
median 146
p90 178
calls · errors 39 · 0
delivery 3.0 s
SR 24 kHz
Smallest.ai Lightning v3.1 Pro
transport http
median 173
p90 220
calls · errors 36 · 1
delivery 1.5 s
SR 24 kHz
ElevenLabs Eleven v4 Turbo
transport http
median 180
p90 214
calls · errors 36 · 0
delivery 2.9 s
SR 24 kHz
Soniox TTS v2
transport http
median 196
p90 276
calls · errors 38 · 0
delivery 16.7 s
SR 24 kHz
Fish Audio Drama 3
transport http
median 205
p90 471
calls · errors 36 · 0
delivery 3.9 s
SR 44.1 kHz
Deepgram Flux TTS
transport http
median 206
p90 277
calls · errors 33 · 1
delivery 12.1 s
SR 24 kHz
StepFun Step-Audio 3 TTS
transport http
median 234
p90 273
calls · errors 25 · 0
delivery 4.3 s
SR 24 kHz
xAI TTS
transport http
median 268
p90 363
calls · errors 36 · 0
delivery 1.9 s
SR 24 kHz
Resemble Ultra
transport http
median 296
p90 496
calls · errors 24 · 0
delivery 1.9 s
SR 48 kHz
Async Flash v1.5
transport http
median 345
p90 369
calls · errors 33 · 1
delivery 4.1 s
SR 24 kHz
Bland Speech v3
transport http
median 354
p90 677
calls · errors 36 · 0
delivery 1.6 s
SR 24 kHz
Giggy TTS
transport http
median 453
p90 862
calls · errors 36 · 0
delivery 0.9 s
SR 24 kHz
ElevenLabs Eleven v4
transport http
median 573
p90 673
calls · errors 36 · 0
delivery 3.5 s
SR 24 kHz
HeyGen Voice 1
transport http
median 623
p90 1,364
calls · errors 36 · 0
delivery 1.7 s
SR 44.1 kHz
Mistral Voxtral TTS 3
transport http
median 653
p90 1,158
calls · errors 36 · 0
delivery 2.6 s
SR 24 kHz
OpenAI gpt-realtime-2.1
transport realtime-ws
median 663
p90 796
calls · errors 39 · 0
delivery 4.5 s
SR 24 kHz
Gemini 3.8 Flash TTS
transport http
median 760
p90 983
calls · errors 35 · 0
delivery 4.8 s
SR 24 kHz
Microsoft MAI-Voice-2.1
transport http
median 817
p90 1,321
calls · errors 36 · 0
delivery 1.2 s
SR 24 kHz
OpenAI gpt-4o-mini-tts
transport http
median 862
p90 1,566
calls · errors 36 · 0
delivery 3.0 s
SR 24 kHz
VUI Labs Luna
transport http
median 1,236
p90 1,303
calls · errors 36 · 0
delivery 2.4 s
SR 24 kHz
MiniMax speech-2.8-hd
transport http
median — †
p90 —
calls · errors 0 · 0
delivery 1.4 s
SR 32 kHz
HTTP vs WebSocket
Vendors offering both, sampled in alternating calls: median ms to first speech (calls/errors).
Vendor
HTTP ms
WebSocket ms
Board uses
Note
Cartesia Sonic 3.6
HTTP 140 (36/0)
WS 125 (36/0)
board uses websocket
Socket clearly faster to first audio; board uses it.
AssemblyAI TTS-1
HTTP 67 (36/0)
WS 30 (36/0)
board uses websocket
Both fast; socket a little faster; board uses it.
Deepgram Flux TTS
HTTP 206 (33/1)
WS 819 (36/0)
board uses http
Socket sends ~0.6 s of silence before speech, so first speech is slower than HTTP; board uses HTTP.
xAI TTS
HTTP 268 (36/0)
WS 229 (36/0)
board uses http
Close; HTTP kept.
Inworld TTS-2
HTTP 98 (39/0)
WS 95 (39/0)
board uses http
Close; HTTP kept.
Soniox TTS v2
HTTP 196 (38/0)
WS 212 (39/0)
board uses http
Socket somewhat faster; both pace audio near playback speed.
Smallest.ai Lightning v3.1 Pro
HTTP 173 (36/1)
WS 161 (36/1)
board uses http
Close on first audio; socket's 'complete' message lags ~4 s.
Murf Falcon 2.1
HTTP 146 (39/0)
WS 143 (39/0)
board uses http
Socket somewhat faster.
StepFun Step-Audio 3 TTS
HTTP 234 (25/0)
WS 240 (1/14)
board uses http
Socket session drops or stalls often; HTTP kept.
Each clip's request, model id, voice id and timings are in its JSON sidecar.
