podcast-tokenized-bg3.5-enj5-with-speaker-embeddings

This dataset extends TTS-AGI/podcast-tokenized-bg3.5-enj5 with speaker embeddings, cosine similarity scores, speaker cluster assignments, and reference-match flags for each sample.

What was added

Each sample's JSON metadata is augmented with the following fields:

Field	Type	Description
target_speaker_embedding	list[float] (128-dim)	L2-normalized speaker embedding of the target audio
ref_speaker_embedding	list[float] (128-dim)	L2-normalized speaker embedding of the reference audio
speaker_cosine_similarity	float	Cosine similarity between target and reference embeddings
ref_matches_target	bool	Whether reference and target are the same speaker (cosine similarity >= 0.42)
speaker_cluster_id	int	Nearest cluster assignment (0-9999) from 10,000 precomputed speaker centroids

The .npy and .ref.npy latent files are byte-identical to the source dataset.

Models used

• DACVAE (facebook/dacvae-watermarked): Decodes latent representations back to 48kHz audio
• Speaker-wavLM-tbr (Orange/Speaker-wavLM-tbr): Extracts 128-dim timbral speaker embeddings from 16kHz audio (WavLM-Large backbone + stats pooling + linear head)

Pipeline

1. Decode DACVAE latents (target and reference) to 48kHz waveforms
2. Truncate to 20s, resample to 16kHz on GPU
3. Extract speaker embeddings (batched, both clips in one forward pass)
4. Compute cosine similarity (dot product of L2-normalized embeddings)
5. Assign ref_matches_target using the F1-optimal threshold (0.42)
6. Assign speaker_cluster_id via nearest centroid lookup against 10,000 precomputed EmoLIA speaker centroids (L2-normalized cosine similarity)
7. Augment JSON metadata, repackage tar, upload

Processed on 8x A100 GPUs in ~34 hours (embeddings), then CPU-only cluster assignment in ~6 hours. 481 tar files, 1,828,199 samples, zero failures.

Dataset statistics

• Total samples: 1,828,199
• Tar files: 481 (up to 4,096 samples each)
• Format: WebDataset tar files, each sample = {key}.npy + {key}.ref.npy + {key}.json

Reference match statistics

Metric	Value
Ref matches target (same speaker)	1,560,471 (85.4%)
Ref does not match target	267,728 (14.6%)
Threshold used	0.42 (F1-optimal)

85.4% of samples have a reference audio clip from the same speaker as the target, making them suitable for voice-cloning and speaker-conditioned TTS tasks without further filtering. The remaining 14.6% have mismatched speakers between reference and target.

Speaker cluster statistics

Samples are assigned to 10,000 speaker centroids derived from EmoLIA speaker embeddings. Each sample's target embedding is matched to its nearest centroid via cosine similarity.

Metric	Value
Centroids available	10,000
Centroids with samples	9,204 (92.0%)
Centroids unused	796 (8.0%)

Cluster size distribution

Statistic	Value
Mean cluster size	198.6
Median cluster size	63
Max cluster size	7,591 (cluster #9997)
Min cluster size	1 (332 singleton clusters)
p10	4
p25	18
p75	195
p90	498
p95	852
p99	2,047
Gini coefficient	0.713

The distribution is heavily right-skewed: 7.6% of clusters (702) account for 50% of all samples, while 39.1% of clusters (3,597) cover 90%. The median cluster has only 63 samples while the mean is 198.6, indicating a long tail of small clusters.

Size range	Clusters	Share
1-9	1,575	17.1%
10-49	2,553	27.7%
50-99	1,426	15.5%
100-199	1,386	15.1%
200-499	1,347	14.6%
500-999	578	6.3%
1,000-1,999	240	2.6%
2,000-4,999	93	1.0%
5,000-9,999	6	0.1%

Balancing strategies

For training with balanced speaker representation, the following capping strategies are available. Each row shows the effect of capping the maximum samples per cluster at a given level.

Target size	Cap per cluster	Samples kept	Clusters capped	Discarded	Upsample needed
100K	11	90,311	7,418	1,737,888	9,689
1M	109	574,441	3,457	1,253,758	425,559
5M	544	1,290,000	837	538,199	3,710,000
10M	1,087	1,566,597	295	261,602	8,433,403

At the 1M target size, capping at 109 samples per cluster retains 574K natural samples and requires upsampling 426K to fill the remaining quota across underrepresented clusters. This provides a practical balance between speaker diversity and data efficiency.

Report files

• cluster_statistics.csv -- Per-cluster sample counts sorted by frequency (9,204 rows)
• cluster_report.html -- Interactive HTML report with distribution visualizations and balancing strategy tables

Speaker similarity threshold analysis

A threshold analysis was conducted using Gemini 2.5 Flash as an independent speaker verification judge. 100 audio pairs were sampled (50 target-reference, 25 same-speaker target-target, 25 different-speaker target-target), decoded, and sent to Gemini for same/different speaker judgment. Results were compared against the cosine similarity scores.

Full analysis PDF: Raw results:

Results

Metric	Value
Same-speaker cosine similarity (mean +/- std)	0.851 +/- 0.181
Different-speaker cosine similarity (mean +/- std)	0.418 +/- 0.408
ROC AUC	0.778
F1-optimal threshold	0.422
Conservative (safe) threshold	0.960

Recommended thresholds

Cosine similarity	Interpretation
>= 0.96	High confidence same speaker
0.42 - 0.96	Uncertain / requires further analysis
< 0.42	High confidence different speaker

Notes

• The F1-optimal threshold (0.422) achieves 74% accuracy and F1 of 0.759
• The separation between classes is 0.73 sigma (modest overlap)
• Gemini agreed with expected labels in 67% of cases -- disagreements were predominantly on target-reference pairs where Gemini detected audio quality/style differences (likely DACVAE reconstruction artifacts) rather than true speaker identity changes
• For filtering out mismatched speakers, cosine < 0.42 is a reliable cutoff