audio-harvest/physics-audio-sim

Physics Audio Sim

English | 繁體中文

English

Sound effects generated by computer models of physics and physics-inspired sound design: things hitting, rolling, breaking, bubbling, blowing, rain, thunder, fire and machines running. They are not recordings. Each row is one WAV file plus the settings used to make it, so you always know what produced the sound.

Release v0.8.1 has 15750 sounds in the main set. They were made with the Sound Design Toolkit (SDT) (6750), Andy Farnell's Designing Sound Pure Data examples (4500), the Synthesis ToolKit (STK) PhISEM models (3825) and the Faust physical modeling library (675).

What is inside

Subset	Sound	Rows	Made with
airflow	Blowing wind	225	SDT
airflow	Whip crack	225	SDT
airflow	Wind blowing across a tube or bottle	225	SDT
airflow	Broadband wind	225	SDT
airflow	Wind whistling past a thin object	225	SDT
deformation	Crumpling or crushing	225	SDT
deformation	Crunching	225	STK PhISEM
electrical	Electrical hum and sparks	225	Farnell (Pure Data)
explosion	Explosion heard from a distance	225	SDT
explosion	Explosion with fireball	225	Farnell (Pure Data)
explosion	Gunfire	225	Farnell (Pure Data)
explosion	Gunfire	225	SDT
fire	Fire burning	225	Farnell (Pure Data)
fluid	A single bubble in water	225	SDT
fluid	Bubbling water	225	Farnell (Pure Data)
fluid	Dripping water	225	SDT
fluid	Water filling a container	225	SDT
fluid	Running or pouring liquid (many bubbles)	225	SDT
fluid	Gushing water	225	SDT
fluid	Pouring water into a container	225	Farnell (Pure Data)
fluid	Water drops	225	STK PhISEM
fluid	Running water	225	Farnell (Pure Data)
friction	Tin can rolling	225	Farnell (Pure Data)
friction	Rubbing with set force and pressure	225	SDT
friction	Creaking door	225	Farnell (Pure Data)
friction	Guiro scraped	225	STK PhISEM
friction	Ratchet wrench	225	STK PhISEM
friction	Ball rolling on a surface	225	SDT
friction	Rubbing back and forth	225	SDT
friction	Rubbing and scraping together	225	SDT
friction	Sandpaper	225	STK PhISEM
friction	Scraping along a surface	225	SDT
impact	Tuned bamboo rattle (angklung)	225	STK PhISEM
impact	Ball bouncing until it stops	225	Farnell (Pure Data)
impact	Bamboo wind chimes	225	STK PhISEM
impact	Object dropped and bouncing	225	SDT
impact	Can with loose bits shaken	225	STK PhISEM
impact	Struck object with set pitch and decay	225	SDT
impact	Djembe	225	Faust physmodels
impact	Something breaking apart	225	SDT
impact	Free metal bar	225	Faust physmodels
impact	Marimba bar	225	Faust physmodels
impact	Two ringing objects hitting each other	225	SDT
impact	Coffee mug rattle	225	STK PhISEM
impact	Big rocks	225	STK PhISEM
impact	Small pebbles	225	STK PhISEM
impact	Cabasa	225	STK PhISEM
impact	Maraca	225	STK PhISEM
impact	Shekere	225	STK PhISEM
impact	Sleigh bells	225	STK PhISEM
impact	Tambourine	225	STK PhISEM
impact	Bundle of sticks	225	STK PhISEM
impact	Single hit	225	SDT
impact	Struck string	225	SDT
impact	Old telephone bell	225	Farnell (Pure Data)
machine	Blender	225	SDT
machine	Car engine heard inside the car	225	SDT
machine	Clock ticking	225	Farnell (Pure Data)
machine	Combustion engine	225	SDT
machine	Electric fan	225	Farnell (Pure Data)
machine	Small electric motor	225	SDT
machine	Refrigerator hum	225	SDT
machine	Helicopter	225	Farnell (Pure Data)
machine	Jet engine	225	Farnell (Pure Data)
machine	Electric motor starting and stopping	225	Farnell (Pure Data)
machine	Photocopier	225	SDT
machine	Piston engine	225	Farnell (Pure Data)
weather	Rain, light to heavy	225	Farnell (Pure Data)
weather	Thunderclap and rumble	225	Farnell (Pure Data)
weather	Outdoor wind	225	Farnell (Pure Data)

How to load

You need pip install "datasets[audio]".

from datasets import load_dataset

ds = load_dataset("audio-harvest/physics-audio-sim", "impact", split="train", revision="v0.8.1")
print(ds[0]["description_en"])

Load one group at a time by its name: airflow, deformation, electrical, explosion, fire, fluid, friction, impact, machine, weather. There is no combined subset; to use everything, load each group and concatenate them. Every subset has train, validation and test splits (60/20/20). They are made inside each sound type: each combination of settings is in exactly one split, and every individual setting value appears in train.

Bounce benchmark (bounce_benchmark)

A separate set for testing whether audio embeddings keep physical information. A ball is dropped and bounces. Only two things change:

• drop height drop_height_m: 15 values, 0.1 to 1.5 m
• coefficient of restitution restitution: 15 values, 0.3 to 0.9

That gives 15 x 15 = 225 sounds. The ball, floor and recording setup are the same in every file, so differences come only from height and restitution. Each text is just the two values, for example "A ball is dropped from 0.5 m; the coefficient of restitution is 0.814."

Split	Rows	Use
train	135	Fit your probe
validation	45	Tune the probe
test	45	Final score only

• Each (height, restitution) pair is in exactly one split. Test pairs are never in train, but every height value and every restitution value appears in train.
• Volume is not adjusted per file here. All 225 files share one gain, so a higher drop is louder, as in reality. The loudest file peaks at -3 dBFS.
• benchmark/bounce_he_v1/test_pairs.jsonl lists 113 test pairs that differ in only one value (51 differ in height, 62 in restitution). Give a model both sounds and both texts and check whether it matches them correctly; guessing gets 50%.
• Files are 12 s long so the longest bounce sequence fits.

from datasets import load_dataset

bench = load_dataset("audio-harvest/physics-audio-sim", "bounce_benchmark", revision="v0.8.1")
print(bench["train"][0]["drop_height_m"], bench["train"][0]["restitution"])

Main columns

Column	Meaning
audio	The sound (48 kHz, mono)
description_en, description_zh	A short sentence describing the settings, in English and Traditional Chinese
split	train, validation or test
phenomenon, event_type	Sound group and type, as in the table above
generation_backend	Which library made it: sdt, farnell_pd, stk_phisem or faust
model	Which model or example inside that library
spectral_centroid_hz	Spectral centroid measured with SDT's analysis tools, for every row (pitch and more are in metadata)
gain_linear	How much the volume was changed (see below)
metadata	All settings and measurements, as JSON text (json.loads(row["metadata"]))

Good to know

• Volume: one gain per sound type. Inside one sound type (one scene_id) loudness is comparable, so a harder hit is louder. Different sound types have different gains, so do not compare loudness across them. To get the raw simulator level, divide the audio by gain_linear.
• The descriptions are generated from the settings. Nobody listened to each file to write them, and the settings are not measured from real objects. For example, a "resonator preset" is not a real wood or metal sample.
• The struck string (string_strike) uses a connection inside SDT that we designed ourselves; SDT has no example for it.
• Sounds with a scene_id starting with sdt_scene_ (dripping, blender, car and so on) rebuild the example patches that ship with SDT. Small helper parts that exist only in Max (timers, ramps, random walks) were rebuilt in Python; metadata.routing names the example each one comes from.
• Rows made with farnell_pd render the Pure Data examples from Andy Farnell's book Designing Sound (rain, fire, engines, helicopters and so on). These are sound designs inspired by physics, not physics simulations. Some controls (for example "ball pitch x") are multipliers we added to the original constants; each row lists every change in metadata.approximations.
• Rows made with stk_phisem use Perry Cook's PhISEM particle model from the Synthesis ToolKit (STK): shakers, rocks, sandpaper, water drops and more. "Object setting" is the model's 0-128 control.
• Rows made with faust use percussion models from the Faust physical modeling library: a marimba bar, a djembe, and an ideal free metal bar whose vibration modes we set from beam theory.

License and citation

Audio and annotations: CC BY 4.0. The generator code keeps its own license and is not part of this one: SDT is GPL-3.0, STK is MIT-style, the Faust libraries are LGPL with an exception for generated code, and Andy Farnell's Pure Data examples may be freely distributed and modified (see ATTRIBUTIONS.md).

AudioHarvest. Physics Audio Sim. Release v0.8.1. https://huggingface.co/datasets/audio-harvest/physics-audio-sim

繁體中文

這是用物理模型和參考物理的音效設計「算」出來的聲音，不是錄音。內容包括撞擊、滾動、碎裂、氣泡、 風聲、雨、雷、火和機器運轉聲。每一筆是一個 WAV 檔，加上產生它的參數。

v0.8.1 版主資料共有 15750 筆。產生方式：Sound Design Toolkit（SDT）（6750 筆）、Andy Farnell《Designing Sound》的 Pure Data 範例（4500 筆）、Synthesis ToolKit（STK） 的 PhISEM 模型（3825 筆）、Faust 物理模型庫（675 筆）。

內容

子集	聲音	筆數	生成方式
airflow	吹氣風聲	225	SDT
airflow	甩鞭聲	225	SDT
airflow	風吹過管口或瓶口	225	SDT
airflow	寬頻風聲	225	SDT
airflow	風吹過細長物體的哨音	225	SDT
deformation	揉皺或壓碎	225	SDT
deformation	壓碎聲	225	STK PhISEM
electrical	電流嗡嗡聲與火花	225	Farnell（Pure Data）
explosion	從遠處聽到的爆炸	225	SDT
explosion	爆炸與火球	225	Farnell（Pure Data）
explosion	槍聲	225	Farnell（Pure Data）
explosion	槍聲	225	SDT
fire	火在燒	225	Farnell（Pure Data）
fluid	水中單顆氣泡	225	SDT
fluid	水中冒泡	225	Farnell（Pure Data）
fluid	滴水	225	SDT
fluid	倒水裝滿容器	225	SDT
fluid	流動或倒水聲（大量氣泡）	225	SDT
fluid	湧水	225	SDT
fluid	把水倒進容器	225	Farnell（Pure Data）
fluid	水滴	225	STK PhISEM
fluid	流水	225	Farnell（Pure Data）
friction	鐵罐滾動	225	Farnell（Pure Data）
friction	可調推力與壓力的摩擦	225	SDT
friction	門的嘎吱聲	225	Farnell（Pure Data）
friction	刮瓜	225	STK PhISEM
friction	棘輪扳手	225	STK PhISEM
friction	球在表面上滾動	225	SDT
friction	來回摩擦	225	SDT
friction	摩擦加刮擦	225	SDT
friction	砂紙	225	STK PhISEM
friction	沿表面刮擦	225	SDT
impact	竹搖琴（昂格隆）	225	STK PhISEM
impact	球彈跳到停住	225	Farnell（Pure Data）
impact	竹風鈴	225	STK PhISEM
impact	物體落下並彈跳	225	SDT
impact	搖晃裝著顆粒的鋁罐	225	STK PhISEM
impact	可調音高與衰減的敲擊	225	SDT
impact	非洲鼓	225	Faust physmodels
impact	物體斷裂碎開	225	SDT
impact	自由金屬棒	225	Faust physmodels
impact	馬林巴琴鍵	225	Faust physmodels
impact	兩個會響的物體互撞	225	SDT
impact	馬克杯碰撞聲	225	STK PhISEM
impact	大石頭	225	STK PhISEM
impact	小石子	225	STK PhISEM
impact	卡巴薩	225	STK PhISEM
impact	沙鈴	225	STK PhISEM
impact	葫蘆珠網搖鈴	225	STK PhISEM
impact	雪橇鈴	225	STK PhISEM
impact	鈴鼓	225	STK PhISEM
impact	一把木棒	225	STK PhISEM
impact	單次敲擊	225	SDT
impact	被敲擊的弦	225	SDT
impact	老式電話鈴	225	Farnell（Pure Data）
machine	果汁機	225	SDT
machine	車內聽到的引擎聲	225	SDT
machine	時鐘滴答	225	Farnell（Pure Data）
machine	內燃引擎	225	SDT
machine	電風扇	225	Farnell（Pure Data）
machine	小型電動馬達	225	SDT
machine	冰箱運轉聲	225	SDT
machine	直升機	225	Farnell（Pure Data）
machine	噴射引擎	225	Farnell（Pure Data）
machine	電動馬達啟動到停下	225	Farnell（Pure Data）
machine	影印機	225	SDT
machine	活塞引擎	225	Farnell（Pure Data）
weather	雨（小雨到大雨）	225	Farnell（Pure Data）
weather	打雷與隆隆聲	225	Farnell（Pure Data）
weather	戶外的風	225	Farnell（Pure Data）

讀取方式

需要先 pip install "datasets[audio]"，程式碼同上方英文段落。 一次讀取一類，名稱為：airflow, deformation, electrical, explosion, fire, fluid, friction, impact, machine, weather。沒有合併全部的子集；要用全部資料，請分別讀取各類再合併。 每個子集都有 train、validation、test（60/20/20），在每一種聲音內各自切分： 每一組參數組合只會出現在一個 split，每一個參數值都會出現在 train。

彈跳 benchmark（bounce_benchmark）

另外一組資料，用來測音訊 embedding 有沒有保留物理資訊。 球落下後連續彈跳，只改兩件事：

• 落下高度 drop_height_m：15 種，0.1 到 1.5 公尺
• 恢復係數 restitution：15 種，0.3 到 0.9

共 15 × 15 = 225 段。球、地面和錄音條件全部相同，所以聲音差異只來自高度和恢復係數。 文字描述只寫這兩個數值，例如「球從 0.5 公尺落下，恢復係數為 0.814。」

Split	筆數	用途
train	135	訓練 probe
validation	45	調整 probe
test	45	只用來算最終分數

• 每個（高度, 恢復係數）組合只會出現在一個 split。test 的組合沒有出現在 train， 但每一個高度值和每一個恢復係數值都有出現在 train。
• 這組資料沒有逐檔調音量。 225 個檔案共用同一個增益，所以落得越高越大聲，和真實情況一樣；最大聲的檔案峰值是 −3 dBFS。
• benchmark/bounce_he_v1/test_pairs.jsonl 列出 113 組只差一個數值的 test 配對 （51 組只差高度、62 組只差恢復係數）。把兩段聲音和兩句描述給模型配對，亂猜的正確率是 50%。
• 每段 12 秒，確保最長的彈跳過程也能完整收錄。

主要欄位

欄位	說明
audio	音檔（48 kHz、單聲道）
description_en、description_zh	依參數產生的一句英文／繁中說明
split	train、validation 或 test
phenomenon、event_type	聲音類別，對應上表
generation_backend	用哪個函式庫做的：sdt、farnell_pd、stk_phisem 或 faust
model	該函式庫裡的哪個模型或範例
spectral_centroid_hz	用 SDT 分析工具量到的頻譜重心（音高等其他特徵在 metadata）
gain_linear	音量調整倍率（見下方）
metadata	完整參數與量測值，存成 JSON 文字（用 json.loads(row["metadata"]) 讀取）

使用前請注意

• 音量：同一種聲音共用一個增益。 同一個 scene_id 之內可以比較音量，例如敲得越用力越大聲； 不同種類的聲音增益不同，不能互相比較。要還原模擬器原始音量，把音訊除以 gain_linear。
• 文字說明是依參數自動產生的，不是人聽完寫的；參數也不是量測真實物體得到的。 例如「共振體 preset」並不代表真的木頭或金屬。
• 被敲擊的弦（string_strike）在 SDT 裡的接法是我們自己設計的，SDT 沒有官方範例。
• scene_id 以 sdt_scene_ 開頭的聲音（滴水、果汁機、車內引擎等），是照 SDT 附帶的範例重建的。 範例裡只有 Max 才有的小元件（計時、漸變、隨機變化）改用 Python 重做；metadata.routing 會寫出來源範例。
• farnell_pd 的聲音來自 Andy Farnell《Designing Sound》書中的 Pure Data 範例（雨、火、引擎、直升機等）。 它們是參考物理設計的音效，不是物理模擬。部分參數（例如「球的音高 ×」）是我們在原始常數上加的倍率； 每一筆的所有改動都寫在 metadata.approximations。
• stk_phisem 的聲音用 Synthesis ToolKit（STK）裡 Perry Cook 的 PhISEM 粒子模型：沙鈴、石頭、砂紙、水滴等。 「物件設定」是模型的 0–128 控制值。
• faust 的聲音用 Faust 物理模型庫的敲擊模型：馬林巴琴鍵、非洲鼓， 以及一根依樑理論自己設定振動模態的理想自由金屬棒。

授權

音檔與標註採 CC BY 4.0。產生聲音的程式各有自己的授權，不在此範圍內：SDT 是 GPL-3.0，STK 是 MIT 類授權， Faust 函式庫是 LGPL（產生的程式碼另有例外），Andy Farnell 的 Pure Data 範例可自由散布與修改（見 ATTRIBUTIONS.md）。