Gametime Benchmark

The Gametime dataset provides lightweight, streaming-friendly splits for TTS/ASR/SpokenLM prototyping.
For full details, please refer to the paper:
👉 Game-Time: Evaluating Temporal Dynamics in Spoken Language Models

---

📦 Download Options

1️⃣ Recommended — Full ZIP Download

If you prefer the original folder layout you can download one of the ZIPs packaged in gametime/download/. There are two kinds available in this repository:

• gametime/download/basic_instructions.zip — unpacks to:

basic_instructions/
├── text/
│   ├── *-dataset.json         # per-dataset JSON manifest(s)
├── audios/
│   ├── <dataset_id>/
│   │   └── test/*.wav
├── alignments/                 # per-audio alignment files 
│   ├── <dataset_id>/
│   │   ├── <stem>.jsonl

• gametime/download/advanced_instructions.zip — unpacks to:

advanced_instructions/
├── text/
│   ├── *-dataset.json         # per-dataset JSON manifest(s) with timing tokens
├── audios/
│   ├── <dataset_id>/
│   │   └── test/*.wav
├── alignments/                 # per-audio alignment files 
│   ├── <dataset_id>/
│   │   ├── <stem>.jsonl

Notes:

• Each ZIP in gametime/download/ preserves the original source tree names (basic_instructions/ or advanced_instructions/).

Download example (Hugging Face):

from huggingface_hub import hf_hub_download
import os

path = hf_hub_download(
    repo_id="gametime-benchmark/gametime",
    repo_type="dataset",
    filename="download/basic_instructions.zip",
    revision="main",
    local_dir=".",
)
print("saved to:", path)

Unzip example:

unzip gametime/download/basic_instructions.zip

---

2️⃣ Optional — Stream from Hugging Face

from datasets import load_dataset

# Load Basic test split
ds_basic = load_dataset("gametime-benchmark/gametime", "basic", split="test", streaming=True)
ex = next(iter(ds_basic))
wav = ex["audio"]["array"]            # numpy float array
sr  = ex["audio"]["sampling_rate"]    # int, e.g. 24000
print(ex["id"], sr, len(wav), ex["text"])

# Load Advanced test split
ds_adv = load_dataset("gametime-benchmark/gametime", "advanced", split="test", streaming=True)
ex_adv = next(iter(ds_adv))
wav_adv = ex_adv["audio"]["array"]
sr_adv  = ex_adv["audio"]["sampling_rate"]
print(ex_adv["id"], sr_adv, len(wav_adv), ex_adv["text"])

• Works with streaming=True — no full download needed
• Audio is auto-decoded via the datasets Audio feature (requires soundfile / libsndfile under the hood)

---

📑 Schema

Each row has the following columns (in order):

Splits: each config (basic, advanced) provides a single test split.

---

📚 Citation

If you use this dataset, please cite:

@article{chang2025gametime,
  title   = {Game-Time: Evaluating Temporal Dynamics in Spoken Language Models},
  author  = {Kai-Wei Chang and En-Pei Hu and Chun-Yi Kuan and Wenze Ren and Wei-Chih Chen and Guan-Ting Lin and Yu Tsao and Shao-Hua Sun and Hung-yi Lee and James Glass},
  year    = {2025},
  journal = {arXiv preprint arXiv:2509.26388},
  url     = {https://arxiv.org/abs/2509.26388}
}

---