Aural One

Native audio in, structured decisions out. Aural One is an early Gemma 4 E2B preview for questions that depend on what a recording actually sounds like. Supply audio, a written state, and named questions; the model scores the options for each question in a single native-audio model. It does not require a speech-to-text or external sound classifier in its inference path.

Model weights and card · Evaluation details · Fine-tuning guide · Training and data · Release checklist

• Scores 2–8 supplied options for each of up to 32 named questions. Binary No/Yes and ordinal decisions use the same interface.
• Accepts written state alongside the audio, so the same recording can answer different questions.
• Uses Gemma 4 E2B's native audio tower plus a frozen adapter and 54.3 million trained acoustic/projection weights. The language model stayed frozen during the final training stage.
• Handles recordings up to 60 seconds in the tested path. The included simple loader runs each question separately; the timing study used a shared-audio serving path.

These are different datasets and scopes, not one combined benchmark score. The complete evaluation table includes other languages, longer natural speech, sound events, calibration, client latency, throughput, and cost. This is an early research release; performance is uneven across sounds and emotions, and the sub-second result above is inside the pod rather than Tokyo-to-server latency.

Use Python 3.12 and an NVIDIA GPU. The tested environment used PyTorch 2.13.0 with CUDA 13.0, Transformers 5.17.0, and PEFT 0.21.0. Install a PyTorch build appropriate for your GPU first, then:

pip install -e .
python scripts/predict.py examples/request.json

Edit examples/request.json to point audio to an absolute path to your own WAV or FLAC file. The first run downloads the pinned Gemma 4 E2B base and the Aural One delta. The base model is not duplicated in this release. The release loader passed smoke checks on a 24 GB Blackwell GPU partition; its observed PyTorch allocation was 9.8 GiB. A lower minimum VRAM has not been established.

The returned probabilities are normalized over the options supplied in that question. They are not calibrated confidence or a free-form numeric regression output. For a binary answer, give options such as ["No", "Yes"]; for an ordinal value, give the score labels as options.

src/aural_one/       pinned loader, choice scoring and acoustic optimizer
scripts/predict.py   JSON command-line example
scripts/train_full_audio.py, build_schedule.py, evaluate.py   training and evaluation
examples/            request shape
docs/EVALUATION.md   measurements and their scopes
docs/FINE_TUNING.md  data schema and reproducible recipe
docs/TRAINING.md     training recipe, data and attribution
release.json         checkpoint identity and hashes

The model weights are on Hugging Face. No training or evaluation audio is redistributed.

Code and Aural One weight deltas: Apache 2.0. The Gemma 4 E2B base is Apache 2.0 and downloaded separately. Training datasets retain their own licenses; see data attribution.