Which AI voice and music models actually do what you ask?
Measured tests of text-to-speech and music generation models. The same inputs for every model, scoring rules published in full, and the audio behind every number so you can check it with your own ears.
Speech models
Text-to-speech from OpenAI, Google, ElevenLabs, Inworld, MiniMax, xAI, Alibaba and open source. Score out of 100, higher is better.
Tests
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?
Music models
Music generators from Google, ElevenLabs, Stability AI, MiniMax, Meta and open source. Score out of 100, higher is better.
Tests
Lyric accuracy
Given lyrics, does the model sing those words?
Genre
Ask for funk, get funk?
Tempo
Ask for 130 BPM, get 130 BPM?
Length
Ask for 30 seconds, get 30 seconds?
No vocals when asked
Ask for an instrumental, get an instrumental?
Clean ending
Does the track end, or just stop?
Speed
How long do you wait for a track?
How this benchmark works
Measured, not rated
Missing words are counted from transcripts. Tempo is measured from the beat. Pitch change is measured in semitones. No "8 out of 10 for naturalness".
Hear every result
Each model page carries clips from the test, including the ones it got wrong and what it actually said.
Raw models, default settings
Each vendor's own API, default voice, one request per item. No retries or clean-up in between that could hide a failure.
Open about its limits
Ranges are shown on every score, one voice per model is a simplification, and AnyToSpeech sells some of these models. All of it is spelled out.
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.