license: cc-by-nc-4.0
task_categories:
- automatic-speech-recognition
- text-to-speech
language:
- en
CC3M spoken captions (image–speech) — DualCodec pre-tokenized
DualCodec (12 Hz) pre-tokenized speech for omni speech–vision MLLM training. Source: Conceptual Captions 3M (filtered); speech synthesized with CosyVoice2.
Contents (train split)
| Total samples | ≈ 500,000 (image–caption–speech triples) |
| Total audio | ≈ 735 hours |
| WebDataset shards | see repo (.tar) |
| Mean / median duration | 5.29s / 4.92s |
| p90 / max duration | 8.08s / 25.67s |
| Length cap | 30s |
Counts/hours are computed by sampling shards (per-shard sample count × #shards); duration stats from a scanned subset.
Duration distribution
| bucket | share |
|---|---|
| 0-5s | 51.3% |
| 5-10s | 44.9% |
| 10-15s | 3.5% |
| 15-20s | 0.2% |
| >20s | 0.0% |
Format (WebDataset .tar)
Each sample shares a key and consists of {key}.sem.npy, {key}.ac.npy, {key}.txt, {key}.jpg:
.sem.npy— DualCodec semantic codes,int16, shape(T,), vocab 16384..ac.npy— DualCodec acoustic codes,int16, shape(7, T), vocab 4096 per codebook..txt— spoken caption text.- Image (
.jpg): RGB 256×256, joined from the source image corpus by sample key.
Frame rate is 12 Hz, so duration in seconds = T / 12.
Length & padding
Codes are stored at their true variable length (no padding baked in). Samples longer than the training grid are handled at load time. During training, sequences are padded to a fixed 240-frame (20 s) grid by appending the DualCodec encoded-silence column (semantic code 3716, with its matching acoustic column) — i.e. padding is applied as code-level silence, not waveform zeros, so the padded region matches the codec's silence distribution. Samples exceeding the grid (> 240 frames) are skipped rather than cropped to preserve audio–text alignment.
Intended use
image→speech (i2s) and speech→image (s2i) alignment SFT. Semantic + acoustic codes reconstruct waveforms via the DualCodec decoder.