ECAPA-TDNN Tajik Speaker Embedding (Saidzoda Lab)

How to get access: please fill in the short form at saidzoda.com/access. We approve requests after we receive the form.

Fine-tune of speechbrain/spkrec-ecapa-voxceleb (Apache-2.0) on Tajik speaker classification (AAM-softmax), for sharper Tajik speaker embeddings in the voice-cloning / speaker-verification pipeline than the stock VoxCeleb-trained model.

Result

Equal-error-rate (EER), full-utterance embeddings, L2-normalized, raw cosine similarity (no mean_var_norm_emb centering — see Methodology below):

Open-set = 10 speakers never seen during training. See results.json in this repo for full trial counts.

Training

• Trunk: ECAPA-TDNN, kept from speechbrain/spkrec-ecapa-voxceleb (embedding_model.ckpt loaded as-is), fine-tuned. Classifier head: freshly initialized (VoxCeleb's 7205-class head is not reusable), sized to the Tajik train-speaker count.
• AAM-softmax: margin=0.2, scale=30.0. epochs=8, batch_size=32, lr=1e-4, 3s random crops.
• Data: 60-narrator premium Tajik diarized speaker bank (44.1/48kHz, resampled to 16kHz on the fly). Speaker labels are ECAPA-clustering pseudo-labels, not human-verified — a known limitation that applies equally to the stock-vs-finetuned comparison above.
• Hardware: 1x H100 (Nvidia Brev).

Files

• embedding_model.ckpt — ECAPA-TDNN trunk weights (speechbrain state_dict), best checkpoint by validation accuracy.
• classifier.ckpt — AAM-softmax classifier head (Tajik train-speaker classes; not needed for embedding extraction, included for completeness/reproducibility).
• ecapa_common.py — the exact model-construction + feature-pipeline code used to build this architecture (needed because this checkpoint is a raw speechbrain state_dict, not a speechbrain.pretrained package with its own hyperparams.yaml).
• label_map.json — {original pseudo speaker id → contiguous training label} for the classifier head's output space.

How to use

import torch
from huggingface_hub import hf_hub_download

# grab ecapa_common.py from this repo (or copy it locally) for the exact architecture
repo_id = "Tohirju/ecapa-tdnn-tajik"
common_path = hf_hub_download(repo_id, "ecapa_common.py")
ckpt_path = hf_hub_download(repo_id, "embedding_model.ckpt")

import importlib.util
spec = importlib.util.spec_from_file_location("ecapa_common", common_path)
ecapa_common = importlib.util.module_from_spec(spec)
spec.loader.exec_module(ecapa_common)

model = ecapa_common.build_embedding_model(device="cpu")
state_dict = torch.load(ckpt_path, map_location="cpu")
model.load_state_dict(state_dict)
model.eval()

compute_features, mean_var_norm = ecapa_common.build_feature_pipeline(device="cpu")
wav = ecapa_common.load_wav_16k("some_tajik_utterance.wav")
emb = ecapa_common.embed_utterance(wav, compute_features, mean_var_norm, model, device="cpu")
# emb: [192] L2-normalizable speaker embedding

Architecture hyperparameters (input_size=80 mels, channels=[1024,1024,1024,1024,3072], lin_neurons=192) match speechbrain/spkrec-ecapa-voxceleb's hyperparams.yaml ("ECAPA big") verbatim — only the weights differ.

License

Apache-2.0 (inherited from the speechbrain base model).

Caveats

• Training-speaker labels are unverified pseudo-labels from an ECAPA-clustering diarization pass, not human-annotated ground truth (see Training section).
• Eval methodology deliberately skips speechbrain's mean_var_norm_emb global-centering step (would need re-fitting to the fine-tuned embedding distribution); this is applied identically to both the stock and fine-tuned model in the comparison above, so the head-to-head delta is fair, but absolute numbers are not directly comparable to speechbrain's own published verify_files pipeline results.

This repo is public but gated (manual approval), per Saidzoda Lab policy.