| --- |
| language: |
| - en |
| task_categories: |
| - audio-to-audio |
| configs: |
| - config_name: complex |
| data_files: |
| - split: val |
| path: complex_val/metadata.parquet |
| - split: test |
| path: complex_test/metadata.parquet |
| - config_name: selective |
| data_files: |
| - split: train |
| path: selective_train/metadata.parquet |
| - split: val |
| path: selective_val/metadata.parquet |
| - split: test |
| path: selective_test/metadata.parquet |
| dataset_info: |
| - features: |
| - name: sample_id |
| dtype: string |
| - name: split |
| dtype: string |
| - name: task |
| dtype: string |
| - name: task_id |
| dtype: int64 |
| - name: instruction |
| dtype: string |
| - name: input_audio_path |
| dtype: string |
| - name: target_audio_path |
| dtype: string |
| - name: use_base |
| dtype: bool |
| - name: source_label |
| dtype: string |
| - name: base_label |
| dtype: 'null' |
| - name: destination_label |
| dtype: string |
| - name: target_label |
| dtype: string |
| - name: specific_time |
| dtype: string |
| - name: condition_type |
| dtype: string |
| - name: condition_type2 |
| dtype: string |
| - name: change_vol |
| dtype: float64 |
| - name: change_Pitch |
| dtype: float64 |
| - name: base_labels |
| sequence: string |
| - name: add_labels |
| sequence: string |
| - name: remove_labels |
| sequence: string |
| - name: edit_start_sample |
| dtype: int64 |
| - name: edit_end_sample |
| dtype: int64 |
| - name: preserve_start_sample |
| dtype: int64 |
| - name: preserve_end_sample |
| dtype: int64 |
| - name: a_start_sample |
| dtype: 'null' |
| - name: a_end_sample |
| dtype: 'null' |
| - name: b_start_sample |
| dtype: 'null' |
| - name: b_end_sample |
| dtype: 'null' |
| - name: swap_layout |
| dtype: 'null' |
| - name: cfg_scale |
| dtype: float64 |
| - name: setting_name |
| dtype: string |
| - name: pitch_change_label |
| sequence: string |
| - name: volume_change_label |
| sequence: string |
| - name: start_time |
| dtype: float64 |
| - name: end_time |
| dtype: float64 |
| - name: feature_audio1_used |
| dtype: bool |
| - name: feature_audio2_used |
| dtype: bool |
| config_name: complex |
| - features: |
| - name: sample_id |
| dtype: string |
| - name: split |
| dtype: string |
| - name: task |
| dtype: string |
| - name: task_id |
| dtype: int64 |
| - name: instruction |
| dtype: string |
| - name: input_audio_path |
| dtype: string |
| - name: target_audio_path |
| dtype: string |
| - name: use_base |
| dtype: bool |
| - name: source_label |
| dtype: string |
| - name: base_label |
| dtype: string |
| - name: destination_label |
| dtype: string |
| - name: target_label |
| dtype: string |
| - name: specific_time |
| dtype: string |
| - name: condition_type |
| dtype: string |
| - name: condition_type2 |
| dtype: string |
| - name: change_vol |
| dtype: float64 |
| - name: change_Pitch |
| dtype: float64 |
| - name: base_labels |
| dtype: string |
| - name: add_labels |
| dtype: string |
| - name: remove_labels |
| dtype: string |
| - name: edit_start_sample |
| dtype: int64 |
| - name: edit_end_sample |
| dtype: int64 |
| - name: preserve_start_sample |
| dtype: int64 |
| - name: preserve_end_sample |
| dtype: int64 |
| - name: a_start_sample |
| dtype: float64 |
| - name: a_end_sample |
| dtype: float64 |
| - name: b_start_sample |
| dtype: float64 |
| - name: b_end_sample |
| dtype: float64 |
| - name: swap_layout |
| dtype: bool |
| config_name: selective |
| --- |
| |
| [](https://gewu-lab.github.io/SteerSound/) [](https://github.com/GeWu-Lab/SteerSound) [](https://huggingface.co/GeWu-Lab/SteerSound) [](https://huggingface.co/datasets/GeWu-Lab/SteerSound) |
|
|
| # SteerSound audio editing data |
|
|
| **Paper:** Steer Unified Audio Editing Toward Reliable Source Selection |
|
|
| Six tasks: Addition (0), Removal (1), Generation (2), Replacement (3), Volume Modification (4), Pitch Modification (5). Instructions are plain English. Model-specific prompt/control prefixes are not part of the data; the SteerSound text encoder inserts them internally during inference/training. |
|
|
| | Public folder | Rows | Rows/task | Paired target audio | |
| |---|---:|---:|---| |
| | complex_val | 6,000 | 1,000 | Yes | |
| | complex_test | 6,000 | 1,000 | Yes | |
| | selective_train | 61,440 | 10,240 | Yes | |
| | selective_val | 6,000 | 1,000 | No | |
| | selective_test | 6,000 | 1,000 | No | |
| |
| Every folder contains `metadata.parquet`, `audio.zip` and `manifest.json` with SHA256 hashes. Audio members preserve their original bytes, member names and directory layout. Audio is 16 kHz mono PCM16, 10 seconds per sample. ZIP paths are stored relative to the archive, such as `audios/val/input/...wav`; changing the public folder name does not change these member paths. Complex val/test and Selective train contain input and reference-target audio. Selective val/test contain input audio only and have empty target paths. Selective val/test are evaluated with task-success rules. No target audio is synthesized or padded into the metadata. |
| |
| ## Continue training |
| |
| **Pretraining and fine-tuning with standard FM:** follow [Source audio and preparation](https://github.com/GeWu-Lab/SteerSound/blob/main/pretrain_and_standard_FM_finetune/dataset/README.md) to prepare the paper's source audio or your own audio–caption data. This shared guide covers source dataset links, formats, processing steps, and expected outputs. Editing examples are constructed online. Use the [standard training guide](https://github.com/GeWu-Lab/SteerSound/tree/main/pretrain_and_standard_FM_finetune) to pretrain or fine-tune, then evaluate on Complex val/test. |
| |
| **STEER post-training:** start from the released [SteerSound checkpoint](https://huggingface.co/GeWu-Lab/SteerSound) and choose one of these settings: |
| |
| 1. **Complex audio editing data:** prepare the paper's source audio or your own audio–caption data with the same [Source audio and preparation](https://github.com/GeWu-Lab/SteerSound/blob/main/pretrain_and_standard_FM_finetune/dataset/README.md) guide. Apply [Emphasis](https://github.com/GeWu-Lab/SteerSound/tree/main/STEER_post_training/emphasis/complex) or [Recovery](https://github.com/GeWu-Lab/SteerSound/tree/main/STEER_post_training/recovery/complex) with online editing-example construction; validate and evaluate on `complex_val` / `complex_test`. |
| 2. **Selective audio editing data:** download the released `selective_train`, `selective_val`, and `selective_test` from the [🤗 dataset](https://huggingface.co/datasets/GeWu-Lab/SteerSound). Follow [Emphasis / Selective](https://github.com/GeWu-Lab/SteerSound/tree/main/STEER_post_training/emphasis/selective) or [Recovery / Selective](https://github.com/GeWu-Lab/SteerSound/tree/main/STEER_post_training/recovery/selective) to post-train on the supplied editing pairs and evaluate on the corresponding splits. |
|
|
| ## Read metadata |
|
|
| ```python |
| from datasets import load_dataset |
| complex_val = load_dataset("GeWu-Lab/SteerSound", "complex", split="val") |
| selective_train = load_dataset("GeWu-Lab/SteerSound", "selective", split="train") |
| ``` |
|
|
| Install `datasets>=3.6.0` to use the metadata-loading examples. These return the parquet metadata, not decoded audio. Complex and Selective have different schemas, so they are separate configs instead of a null-padded merged table. Within Selective, training-only and evaluation-only fields have a declared common nullable schema so the three splits load together. Existing missing values stay missing; no metadata values or audio are filled in. The original `split` values remain train/val/test; the top-level names identify the dataset and split together. |
|
|
| ## Read a WAV and edit it |
|
|
| ```python |
| import io, zipfile |
| import pandas as pd |
| import soundfile as sf |
| from huggingface_hub import hf_hub_download |
| from steersound import SteerSoundPipeline # install code from the linked GitHub repository |
| |
| repo = "GeWu-Lab/SteerSound" |
| folder = "selective_val" |
| metadata = hf_hub_download(repo, f"{folder}/metadata.parquet", repo_type="dataset") |
| archive = hf_hub_download(repo, f"{folder}/audio.zip", repo_type="dataset") |
| row = pd.read_parquet(metadata).iloc[0] |
| with zipfile.ZipFile(archive) as z: |
| wave, sr = sf.read(io.BytesIO(z.read(row.input_audio_path)), dtype="float32") |
| pipe = SteerSoundPipeline.from_pretrained(repo, device="cuda") |
| # Exact numeric metadata matters: instruction text may round the amount to one decimal. |
| edited = pipe.edit_record(row.instruction, wave, sample_rate=sr, |
| task_id=int(row.task_id), change_vol=float(row.change_vol), |
| change_pitch=float(row.change_Pitch), sample_id=str(row.sample_id)) |
| sf.write("edited.wav", edited, 16000, subtype="PCM_16") |
| ``` |
|
|
| For interactive use, `pipe("Remove the sound of a dog barking.", audio="input.wav")` parses task and amount automatically. See the [model card](https://huggingface.co/GeWu-Lab/SteerSound) and [GitHub inference instructions](https://github.com/GeWu-Lab/SteerSound) for GPU installation, all six instruction examples and estimated VRAM. |
|
|
| ## Metadata and evaluation |
|
|
| `sample_id`, sample order, task IDs, all numeric changes and task-region boundaries match the source data. The model-specific instruction prefix was removed. Complex val/test retain their original reference targets and target paths; Selective val/test have empty target paths. All remaining input WAV bytes are preserved. `input_audio_path` and optional `target_audio_path` name WAV members inside audio.zip. `change_vol` is signed dB; `change_Pitch` is signed semitones; region columns ending in `_sample` use 16 kHz sample coordinates. Source labels, base/add/remove label lists, condition and timing metadata are retained. Do not infer exact changes solely from rounded instruction strings. |
|
|
| The full Complex evaluation protocol uses task-aware CLAP, softmax KL, pitch/volume MAE, LSD and TBG. The released Complex val/test splits include the ground-truth targets required for KL, MAE, LSD, and the full six-task TBG. TBG additionally requires a baseline-model summary on the same split; see the [evaluation guide](https://github.com/GeWu-Lab/SteerSound/tree/main/inference#evaluate). Task-aware CLAP uses labels and generated audio without paired targets. Selective uses the original task-success and error rules; unmeasurable pitch pairs are excluded from the accuracy denominator. Validation selects checkpoints. Test is used only after selection. |
|
|
