Speech Model Benchmark
The same 262 requests sent to every text-to-speech model. Scores are counted from the audio: words missed, numbers misread, pitch that did or did not move.
Leaders
The top model on the measures people ask about most.
Overall ranking
Score out of 100, higher is better. Weighted: Words 25%, Numbers 20%, Long 15%, Repeats 10%, Languages 15%, Emotion 15%.
Models within about 2 points of each other are a tie: the difference is smaller than the measurement noise. Speed and reliability are reported separately and do not affect this score. Rows marked "ours" are AnyToSpeech itself, measured with the same harness and rules.
Every model on every test
Darker is better. Click a column heading to sort by that test, or a test name to open its page.
| Model | Overall | Words | Numbers | Long | Repeats | Languages | Emotion | Reliable | Speed | |
|---|---|---|---|---|---|---|---|---|---|---|
| 1 | Gemini 3.8 Flash-Lite TTSGoogle | 99.2 | 100 | 98.3 | 99.9 | 100 | 99.7 | 97.5 | 100 | 1.2× |
| 2 | AnyToSpeech Standard voicesour own product | 98.9 | 100 | 98.3 | 99.9 | 100 | 100 | 95 | 100 | 1.1× |
| 3 | Gemini 3.8 Flash TTSGoogle | 98.1 | 100 | 98.3 | 99.9 | 100 | 99.7 | 90 | 100 | 1× |
| 4 | AnyToSpeech Premium and Charactersour own product | 93.6 | 100 | 98.3 | 100 | 100 | 99.9 | 60 | 100 | 3.1× |
| 5 | MiniMax Speech 2.8 HDMiniMax | 90.3 | 99.9 | 96.7 | 79.9 | 100 | 99.7 | 60 | 100 | 2.7× |
| 6 | MiniMax Speech 2.8 TurboMiniMax | 89.2 | 99.9 | 93.3 | 79.9 | 100 | 99.6 | 57.5 | 98.1 | 2.9× |
| 7 | Grok TTSxAI | 88.1 | 100 | 96.7 | 99.8 | 100 | 100 | 25 | 100 | 5.4× |
| 8 | Inworld TTS-2Inworld | 84.8 | 100 | 100 | 40 | 99.6 | 98.9 | 60 | 100 | 3.7× |
| 9 | Eleven v3ElevenLabs | 82.3 | 100 | 100 | 60 | 100 | 99.7 | 22.5 | 100 | 2.5× |
| 10 | GPT-4o mini TTSOpenAI | 81.6 | 99.9 | 98.3 | 79.8 | 97.3 | 99.5 | 2.5 | 100 | 4.3× |
| 11 | Kokoro 82MHexgrad | 81.6 | 100 | 83.3 | 100 | 100 | 99.6 | 0 | 100 | 21.4× |
| 12 | Eleven Flash v2.5ElevenLabs | 81.3 | 99.9 | 76.7 | 100 | 100 | 98.8 | 7.5 | 100 | 15.2× |
| 13 | Eleven Multilingual v2ElevenLabs | 79 | 100 | 76.7 | 79.9 | 100 | 98.7 | 12.5 | 100 | 11.6× |
| 14 | Inworld TTS-1.5 MaxInworld | 77.1 | 100 | 100 | 40 | 100 | 97.2 | 10 | 100 | 4.2× |
| 15 | Gemini 2.5 Pro TTSGoogle | 76.8 | 100 | 50 | 59.9 | 100 | 99.9 | 52.5 | 86.8 | 0.9× |
| 16 | Qwen3 TTSAlibaba | 76.3 | 99.9 | 81.7 | 58.6 | 93.8 | 99.9 | 12.5 | 100 | 1.1× |
| 17 | TTS-1OpenAI | 76.2 | 100 | 88.3 | 59.9 | 99.6 | 97.5 | 0 | 100 | 2.8× |
| 18 | TTS-1 HDOpenAI | 70.9 | 100 | 90 | 60 | 94.2 | 63.5 | 0 | 99.6 | 1.9× |
| 19 | Chatterbox MultilingualResemble AI | 59.9 | 100 | 60 | 13.3 | 98.7 | 68.6 | 5 | 100 | 1.7× |
Accuracy against speed
Up is more accurate, right is faster. Hover a point for its name.
The two kinds of point are timed differently, so compare speed within a kind, not across.
Compare models
Pick up to three. Each spoke is one test, scored 0 to 100.
What stands out
Plain sentences are solved
That many models read 60 everyday sentences with 99% of words or more correct. Ordinary prose no longer separates models.
Numbers and symbols still do
The spread on prices, dates, phone numbers and abbreviations, from Gemini 2.5 Pro TTS to Inworld TTS-1.5 Max.
Emotion settings that change nothing
Models where asking for angry, excited, scared or a whisper produced almost no measurable change in the voice.
Requests that returned nothing
Gemini 2.5 Pro TTS had the most failed requests. 0 other models failed at least once. See reliability.
The tests
Each has its own page with the ranking, the rule, and audio.
Word accuracy
Does the voice say every word, and only the words, in ordinary sentences?
Numbers, dates and symbols
Does it read prices, dates, phone numbers, units, emails and abbreviations the way a person would?
Long text
How much text can it take in one request and still read every word?
Repeated words
When the text repeats itself, does the voice repeat it too, or tidy it away?
Other languages
Does the same model stay accurate outside English?
Emotion control
When you ask for a delivery, does the voice measurably change?
Reliability
How often does a request come back with usable audio on the first try?
Speed
How fast is audio produced, compared with its own length?
Think a score is wrong? Check it.
Every test item, scoring rule and known limit is published, and the audio behind each number is on the page.