README.md

metadata

license: mit
datasets:
  - mozilla-foundation/common_voice_17_0
language:
  - en
  - es
  - ar
  - fr
  - de
  - it
  - pt
  - ru
  - zh
  - ja
metrics:
  - accuracy
base_model:
  - hubertsiuzdak/snac_24khz
pipeline_tag: audio-classification
tags:
  - audio
  - language
  - classification

Audio Language Classifier (SNAC backbone, Common Voice 17.0)

Summary:

• Task: Spoken language identification (10 languages)
• Backbone: SNAC (hubertsiuzdak/snac_24khz) with attention pooling
• Dataset: Mozilla Common Voice 17.0 (streaming)
• Sample rate: 24 kHz; Max audio length: 10 s (pad/trim)
• Mixed precision: FP16
• Best validation accuracy: 0.5016
• Test accuracy: 0.3830

Supported languages (labels):

• en, es, fr, de, it, pt, ru, zh-CN, ja, ar

Intended use:

• Classify the language of short speech segments (≤10 s).
• Not for ASR or dialect/variant classification.

Out-of-scope:

• Very long audio, code-switching, overlapping speakers, noisy or music-heavy inputs.

Data:

• Source: Mozilla Common Voice 17.0 (streaming; per-language subset).
• License: CC-0 (check dataset card for details).
• Splits: Official validation/test splits used (use_official_splits: true).
• Optional percent slice per split used during training: 25%.

Model architecture:

• Backbone: SNAC encoder (pretrained).
• Pooling: Attention pooling over time.
• Head:
  • Linear(feature_dim → 512), ReLU, Dropout(0.1)
  • Linear(512 → 256), ReLU, Dropout(0.1)
  • Linear(256 → 10)
• Selective tuning:
  • Start frozen (backbone_tune_strategy: "frozen")
  • Unfreeze strategy at epoch 5: "last_n_blocks" with last_n_blocks: 1
  • Gradient checkpointing enabled for backbone.

Training setup:

• Batch size: 48
• Epochs: up to 100 (early stopping patience: 15)
• Streaming steps per epoch: 500
• Optimizer: AdamW (betas: 0.9, 0.999; eps: 1e-8)
• Learning rate: head 1e-4; backbone 2e-5 (after unfreeze)
• Scheduler: cosine with warmup (num_warmup_steps: 2000)
• Label smoothing: 0.1
• Max grad norm: 1.0
• Seed: 42
• Hardware: CUDA if available; FP16 enabled

Preprocessing:

• Mono waveform at 24 kHz; pad/trim to 10 s.
• Normalization handled by torchaudio/Tensor transforms in pipeline.

Evaluation results:

• Validation:
  • Best accuracy: 0.5016
• Test:
  • accuracy: 0.3830
  • f1_micro: 0.3830
  • f1_macro: 0.3624
  • f1_weighted: 0.3666
  • loss: 2.2467

Files and checkpoints:

• Checkpoints dir: ./training
  • best_model.pt
  • language_mapping.txt (idx: language)
  • final_results.txt

How to use (inference):

import torch
from models import LanguageClassifier

device = "cuda" if torch.cuda.is_available() else "cpu"

# Build and load from a directory containing best_model.pt and language_mapping.txt
model = LanguageClassifier.from_pretrained("training", device=device)

# Single-file prediction (auto resample to 24k, pad/trim to 10s)
label, prob = model.predict("example.wav", max_length_seconds=10.0, top_k=1)
print(label, prob)

# Top-3
top3 = model.predict("example.wav", top_k=3)
print(top3)  # [('en', 0.62), ('de', 0.21), ('fr', 0.08)]

# If you already have a waveform tensor:
#   wav: torch.Tensor [T] at 24kHz (or provide sample_rate to auto-resample)
#   model.predict handles [T] or [B,T]
# label, prob = model.predict(wav, sample_rate=orig_sr, top_k=1)

Limitations and risks:

• Accuracy varies across speakers, accents, microphones, and noise conditions.
• May misclassify short utterances or code-switched speech.
• Not suitable for sensitive decision making without human review.

Reproducibility:

• Default config: ./config.yaml
• Training script: ./train.py
• To visualize internals: CHECKPOINT_DIR=training/ python viualization.py

Citation and acknowledgements:

• SNAC: hubertsiuzdak/snac_24khz
• Dataset: Mozilla Common Voice 17.0