CekuraBench

TTS Model Benchmarks

LeaderboardListenHow we tested

Reads it right or fast

Higher and further left is better · 8 voices that can't take streamed text are not shown

Best quadrantBest picks
Accuracy (%)
Streaming latency, median (ms)

Text-to-speech models read the same agent sentences. This is who starts first and sounds right.

Last updated Methodology by Dileep Chagam

Leaderboard

Pick the measure your agent can't do without.

Streaming latencyFrom the last word · lower is better
#ModelResult
    • Word-by-word text: speech starts 64 ms before the last word arrives.
    • No stop message: the agent drops the stream.
    Gradium TTS: what it's better at
    Non-streaming
    76 ms
    Streaming
    −64 ms
    Stopping
    n/a
    Accuracy
    79.5%
    Naturalness
    4.51
    Price / 1M chars
    $36.00

    voice Everly · streaming

    • Lowest non-streaming latency, 46 ms.
    • One sentence in ten waits 73 ms or more.
    • No stop message: the agent drops the stream.
    Inworld TTS-2 Flash: what it's better at
    Non-streaming
    46 ms
    Streaming
    45 ms
    Stopping
    n/a
    Accuracy
    93.2%
    Naturalness
    4.46
    Price / 1M chars
    $15.00

    voice Sarah · streaming

    • Word-by-word text starts sooner: 69 ms.
    • Audio falls behind playback on 5% of sentences.
    • Reads the text it was sent before its stop takes effect; the agent drops that audio.
    ElevenLabs v4 Turbo: what it's better at
    Non-streaming
    176 ms
    Streaming
    69 ms
    Stopping
    n/a
    Accuracy
    79.9%
    Naturalness
    4.42
    Price / 1M chars
    $40.00

    voice Rachel · streaming

    • Opens with 223 ms of silence.
    • One sentence in ten waits 449 ms or more.
    • Word-by-word text starts sooner: 126 ms.
    • Goes quiet at once when stopped.
    Deepgram Aura-2: what it's better at
    Non-streaming
    281 ms
    Streaming
    126 ms
    Stopping
    40 ms
    Accuracy
    77.2%
    Naturalness
    4.43
    Price / 1M chars
    $30.00

    voice Thalia · streaming

    • No stop message: the agent drops the stream.
    Inworld TTS-2: what it's better at
    Non-streaming
    126 ms
    Streaming
    127 ms
    Stopping
    n/a
    Accuracy
    93.4%
    Naturalness
    4.52
    Price / 1M chars
    $25.00

    voice Sarah · streaming

    • Goes quiet at once when stopped.
    Deepgram Flux: what it's better at
    Non-streaming
    194 ms
    Streaming
    154 ms
    Stopping
    80 ms
    Accuracy
    77.4%
    Naturalness
    4.16
    Price / 1M chars
    $45.00

    voice Haley · streaming

    • Word-by-word text starts sooner: 157 ms.
    • Goes quiet at once when stopped.
    Grok TTS: what it's better at
    Non-streaming
    358 ms
    Streaming
    157 ms
    Stopping
    0 ms
    Accuracy
    87.4%
    Naturalness
    4.17
    Price / 1M chars
    $15.00

    voice carina · streaming

    • Reads the text it was sent before its stop takes effect; the agent drops that audio.
    ElevenLabs Flash v2.5: what it's better at
    Non-streaming
    182 ms
    Streaming
    177 ms
    Stopping
    n/a
    Accuracy
    64.4%
    Naturalness
    4.33
    Price / 1M chars
    $40.00

    voice Rachel · streaming

    • Lowest list price, $14.81 per million characters.
    • Goes quiet at once when stopped.
    Soniox TTS RT v2: what it's better at
    Non-streaming
    241 ms
    Streaming
    183 ms
    Stopping
    0 ms
    Accuracy
    83.6%
    Naturalness
    4.37
    Price / 1M chars
    $14.81

    voice Emma · streaming

    • Opens with 132 ms of silence.
    Cartesia Sonic 3.5: what it's better at
    Non-streaming
    261 ms
    Streaming
    244 ms
    Stopping
    809 ms
    Accuracy
    73.1%
    Naturalness
    4.41
    Price / 1M chars
    $39.20

    voice Skylar · streaming

    • Opens with 124 ms of silence.
    • No stop message: the agent drops the stream.
    Rime Coda: what it's better at
    Non-streaming
    282 ms
    Streaming
    269 ms
    Stopping
    n/a
    Accuracy
    76.7%
    Naturalness
    4.42
    Price / 1M chars
    $50.00

    voice Astra · streaming

    • Opens with 142 ms of silence.
    Cartesia Sonic 3.6: what it's better at
    Non-streaming
    286 ms
    Streaming
    270 ms
    Stopping
    764 ms
    Accuracy
    74.9%
    Naturalness
    4.41
    Price / 1M chars
    $39.20

    voice Skylar · streaming

    • Opens with 125 ms of silence.
    • No stop message: the agent drops the stream.
    Rime Mist v3: what it's better at
    Non-streaming
    292 ms
    Streaming
    293 ms
    Stopping
    n/a
    Accuracy
    77.2%
    Naturalness
    4.34
    Price / 1M chars
    $30.00

    voice Astra · streaming

    • Sounds the most natural to the listener model, 4.54.
    • No stop message: the agent drops the stream.
    Fish Audio Drama 3: what it's better at
    Non-streaming
    321 ms
    Streaming
    312 ms
    Stopping
    n/a
    Accuracy
    91.3%
    Naturalness
    4.54
    Price / 1M chars
    n/a

    voice Delia · streaming

    • No stop message: the agent drops the stream.
    Fish Audio S2.1-Pro: what it's better at
    Non-streaming
    333 ms
    Streaming
    321 ms
    Stopping
    n/a
    Accuracy
    92.0%
    Naturalness
    4.40
    Price / 1M chars
    $15.00

    voice Sarah · streaming

    • Opens with 141 ms of silence.
    • One sentence in ten waits 679 ms or more.
    • No stop message: the agent drops the stream.
    Smallest Lightning v3.1: what it's better at
    Non-streaming
    357 ms
    Streaming
    363 ms
    Stopping
    n/a
    Accuracy
    81.3%
    Naturalness
    4.49
    Price / 1M chars
    $19.50

    voice kelsey · streaming

    • Audio falls behind playback on 7% of sentences.
    • Reads the text it was sent before its stop takes effect; the agent drops that audio.
    ElevenLabs v4: what it's better at
    Non-streaming
    678 ms
    Streaming
    561 ms
    Stopping
    n/a
    Accuracy
    80.4%
    Naturalness
    4.42
    Price / 1M chars
    $80.00

    voice Rachel · streaming

    • No stop message: the agent drops the stream.
    Hume Octave 2: what it's better at
    Non-streaming
    582 ms
    Streaming
    573 ms
    Stopping
    n/a
    Accuracy
    87.4%
    Naturalness
    4.38
    Price / 1M chars
    $100.00

    voice Ava Song · streaming

No figure: ElevenLabs v3, GPT-4o mini TTS, Gemini 3.8 Flash, Gemini 3.8 Flash-Lite, Gemini 3.1 Flash, Qwen3-TTS, MAI-Voice-2.1-Flash, MAI-Voice-2.1.

* Most voices start before the last word only on some sentences; a minus means this voice did on most of them.

Open a row to see what stands out about that voice. Accuracy and naturalness are estimates.

Where the voices part ways

The usual wait is only part of it. Voices also differ on their slow moments, on streamed text, on stopping, and on hard words.

Share of each kind of hard part read right. Darker blue is better. An estimate: speech-to-text transcribers scored it, not people.

ModelNumbers and shorthand7 a.m. · 20% · 36°F · approx.Codes and addressesSAVE20 · help@north-clinic.org · 1-800-555-0199Read from contextSt. Mary's on Main St. · Dr. OkaforMedical termslevothyroxine · gastroenterology
Inworld TTS-2
93%
86%
100%
99%
Inworld TTS-2 Flash
93%
86%
98%
99%
Fish Audio S2.1-Pro
90%
86%
100%
100%
Fish Audio Drama 3
88%
87%
100%
100%
Gemini 3.8 Flash-Lite
85%
85%
98%
99%
Gemini 3.8 Flash
84%
82%
98%
100%
Gemini 3.1 Flash
83%
85%
98%
100%
Hume Octave 2
87%
71%
100%
99%
Grok TTS
82%
86%
100%
99%
ElevenLabs v3
81%
89%
94%
100%
Soniox TTS RT v2
78%
86%
88%
97%
GPT-4o mini TTS
78%
78%
96%
94%
MAI-Voice-2.1
85%
67%
73%
99%
Smallest Lightning v3.1
78%
81%
75%
97%
ElevenLabs v4
72%
83%
88%
100%
ElevenLabs v4 Turbo
71%
82%
88%
100%
Gradium TTS
77%
77%
67%
99%
MAI-Voice-2.1-Flash
79%
67%
75%
92%
Deepgram Flux
72%
81%
69%
97%
Deepgram Aura-2
71%
87%
75%
86%
Rime Mist v3
71%
77%
71%
100%
Rime Coda
70%
79%
81%
93%
Cartesia Sonic 3.6
71%
75%
58%
100%
Cartesia Sonic 3.5
69%
75%
54%
97%
Qwen3-TTS
69%
49%
75%
100%
ElevenLabs Flash v2.5
55%
55%
75%
100%

Hear every model read the same line

Pick a sentence and play each voice, or play them all in turn.

A greeting · 1 of 19

“Thanks for calling Cedar Valley Family Practice, this is Riley. How can I help you today?”

The first read of each sentence, unedited. Every clip plays at the same loudness. All voices are AI-generated.

Put two voices head to head

Pick any two voices and see them side by side. Every pair has its own page with the full numbers and both voices to hear.

Compare
MeasureInworld TTS-2 FlashGradium TTS
  1. Non-streaming46 ms76 ms
  2. Streaming45 ms−64 ms
  3. StoppingNo stop messageNo stop message
  4. Accuracy93.2%79.5%
  5. Naturalness4.464.51
  6. Price / 1M$15.00$36.00
  7. Measures won3 of 62 of 6
The check marks the better value on each measure. Every pair has its own page with the full numbers.Open this pair · 325 pairs in all

The fastest voice against the next six

How we tested

Every model read the same 104 agent sentences, three times each, from one cloud region, timed from the first text sent to the first sound.

26
models
104
sentences
3
reads of each sentence
30,564
requests
Harness and sentences

Full methodology

Step by step
  1. 1

    Same sentences, same place

    104 sentences a phone agent says, from greetings to 500-character replies, each read three times. Every model is called from one cloud region in the US East.

  2. 2

    The clock starts at the text

    Non-streaming latency runs from sending the whole sentence to the first audible sound; streaming latency runs from the last word sent. Connecting is left out; silence at the start is counted.

  3. 3

    Tested the way agents use them

    Whole sentences, text sent word by word, and a stop mid-sentence as when a caller interrupts.

  4. 4

    Measured, or marked as an estimate

    Timings, stopping, gaps and failures come straight from the recordings, and prices from each vendor's page. Two estimates are marked as such: naturalness is UTMOS, an open model that predicts a listening-test score, and reading accuracy is scored by speech-to-text, not by people.

Text-to-speech benchmark FAQ

  • Gradium TTS from Gradium had the lowest median streaming latency, −64 ms, measured from the last word sent to the first audible sound; a minus means it started speaking before the last word on most sentences.

  • TTS RT v2 from Soniox has the lowest list price, $14.81 per million characters. Prices are each vendor's published rate on the date shown with it.

  • TTS-2 scored highest on hard parts, 93.4%: phone numbers, prices, dates, codes, abbreviations and medical terms, as speech-to-text heard them. Inworld TTS-2 Flash, Fish Audio S2.1-Pro are within two points, so treat the top as a tie.

  • Each recording is transcribed by Deepgram Nova-3 and OpenAI Whisper, and a hard part passes when the transcript holds an accepted reading of it; the two are counted together. No person checks the result, so it is an estimate: on one model the two transcribers differ by 10 points. Hover a figure in the leaderboard for each one's score, or hear the recordings in the Listen section.

  • Streaming and non-streaming latency, slow sentences, how much audio a voice still sends after a stop, playback gaps, failed requests and list price, plus estimates of how natural each voice sounds and how accurately it reads numbers, codes and terms.

  • Yes. Every model reads the same 104 sentences three times, called from one cloud region in the US East over its own streaming protocol, with one stock voice.

  • Yes. The Listen section plays every model reading the same sentences, converted to MP3 and otherwise unedited. Playback matches every clip to the same loudness.