Dodona

A voice-native decision model.

Ask your audio about what was said, how it was said, and who said it. Every answer comes with a probability you can trust.

Try it yourself Record your voice or pick an example, type your own questions. No install, free GPU.

Model on Hugging Face · Code on GitHub

real output · a 2.7-second clip of an actor speaking angrily (CREMA-D; 10 of 10 listeners heard anger)
Each percentage is how likely that answer is, not a share of the recording.
yes / no Does the speaker sound angry? Yes 98%
one of six · the top three options Which emotion is the speaker expressing? angry97% happy1% disgust1%
scale · your own question, not verified How upset is the speaker? 7%calm 59%annoyed 34%very upset

What makes it different

One pass, no text to parse The usual way to question an audio language model is to let it write an answer, token by token, for each question, then parse the text. Dodona reads the probability of every answer option straight from the model, for all questions at once: 40 questions take about four times as long as one, not forty times (53 questions on a 10-second window in 1.4 s on one NVIDIA L40S).

Probabilities that mean what they say On the questions it was trained for, when Dodona says 80% it is right about 80% of the time on recordings it never saw (mean calibration error 0.024). The model card shows the full reliability tables, not just one number.

How, who and what, together One model hears how someone feels (emotion, tone), who is speaking (several people or one, voices, laughter) and whether a word was said, so you can ask about all of them in the same pass. For topics, a transcript is still more accurate, and the model card says so.

Ears for Claude

Dodona ships an MCP server, so Claude can ask it questions about an audio file: Claude hands over a recording and its questions, gets calibrated answers back, and reasons over them. The server runs on your own machine and needs a CUDA GPU (about 24 GB); a hosted version that needs no GPU is planned.

Why Dodona

Dodona was the oldest oracle of ancient Greece. Its priests heard the answers in sound: the rustling of a sacred oak, the calls of doves, the ringing of bronze cauldrons in the wind. People wrote their questions on small lead tablets, many of them yes-or-no ("is it better for me to…?"), and many survive. A written question, a recording to listen to, a yes or a no: that is still the whole idea.

Status

Version 1. Open source (Apache-2.0) on GitHub, weights on Hugging Face, with a model card that states what it can and cannot answer.
Two small adapters (about 135 MB) on Qwen2-Audio-7B-Instruct, trained on English speech: AMI meetings, CREMA-D, Common Voice.
Built by one person on one cloud GPU, with every experiment pre-registered.
Limits, stated up front: emotion was learned from actors, some scale questions are weak, and it needs a GPU with about 24 GB.
Not a chat model: Dodona answers questions about audio; it does not hold a conversation.

Contact

Questions, ideas or a use case you would like Dodona to hear: dobaybaha3@gmail.com