README.md

---
language:
- multilingual
tags:
- audio
- text
- multimodal
- seamless
- subtitle-editing-time-prediction
- translation-aware
- language-pairs
library_name: transformers
base_model: facebook/hf-seamless-m4t-medium
license: cc-by-nc-4.0
---

# videoloc/seamless-langpairs

## Model Description

This is a **SeamlessLanguagePairs** model that processes audio and text inputs with both translation awareness and language pair embeddings to predict **Time To Edit (TTE)** for subtitle segments. Given an audio segment and its corresponding subtitle text, the model predicts how much time (in seconds) would be required to edit/refine that subtitle segment, taking into account both whether the subtitle is translated and the specific language pair involved.

The model extends the SeamlessM4T architecture with both translation features and language pair embeddings, providing the most granular control for multilingual scenarios across **5 languages: English, French, Spanish, Italian, and German** with **21 different translation pairs** between them (e.g., EN→FR, ES→DE, IT→EN, etc.).

### Key Features

- **Language Pair Embeddings**: Fine-grained control for 21 language pairs plus "other"
- **Translation-Aware Processing**: Distinguishes between original and translated content
- **Multimodal Processing**: Simultaneously processes audio (16kHz) and text inputs
- **Frozen Encoders**: Uses pre-trained SeamlessM4T encoders (frozen for stability)
- **Enhanced Architecture**: Adds both translation and language pair embeddings
- **TTE Prediction**: Predicts editing time required for subtitle segments
- **Direct Output**: Raw time values in seconds for immediate use

## Model Architecture

The model extends the basic SeamlessM4T architecture with both translation and language pair awareness:

1. **Audio Processing**: 
   - SeamlessM4T speech encoder (frozen) processes raw audio input
   - Audio projection layer maps speech encoder output to 1024 dimensions
   - Mean pooling over sequence length to get fixed-size audio embedding

2. **Text Processing**:
   - SeamlessM4T text encoder (frozen) processes tokenized text input  
   - Text projection layer maps text encoder output to 1024 dimensions
   - Mean pooling over sequence length to get fixed-size text embedding

3. **Translation Feature Processing**:
   - Binary translation flag (0/1) indicating original vs translated content
   - Translation projection layer maps binary input to 32 dimensions
   - Learned embedding helps model distinguish translation effects

4. **Language Pair Processing**:
   - Categorical language pair ID (0-20) for specific language combinations
   - Language pair embedding layer maps IDs to 64-dimensional vectors
   - Captures language-specific temporal alignment patterns

5. **Feature Fusion**:
   - Audio, text, translation, and language pair embeddings are concatenated (2144 total dimensions)
   - Simple concatenation without complex cross-modal interactions

6. **Regression Head**:
   - Multi-layer perceptron: 2144 → 1024 → 512 → 256 → 1
   - ReLU activations and dropout for regularization
   - Single output for TTE prediction (regression, in seconds)

## Quick Start

### Installation
```bash
pip install transformers torch torchaudio huggingface_hub
```

### Basic Usage
```python
from transformers import AutoModel, AutoConfig
from huggingface_hub import hf_hub_download
import torch
import numpy as np
import importlib.util

# Load model - custom architecture requires importing the model class
model_files = hf_hub_download(repo_id="videoloc/seamless-langpairs", filename="modeling_seamless_langpairs.py")
spec = importlib.util.spec_from_file_location("modeling_seamless_langpairs", model_files)
modeling_module = importlib.util.module_from_spec(spec)
spec.loader.exec_module(modeling_module)

# Now load the model using the custom class
config = modeling_module.SeamlessLanguagePairsConfig.from_pretrained("videoloc/seamless-langpairs")
model = modeling_module.HFSeamlessLanguagePairs.from_pretrained("videoloc/seamless-langpairs")

# Load the data collator (included in this repo)
collator_file = hf_hub_download(repo_id="videoloc/seamless-langpairs", filename="data_collator.py")
spec = importlib.util.spec_from_file_location("data_collator", collator_file)
collator_module = importlib.util.module_from_spec(spec)
spec.loader.exec_module(collator_module)

# Initialize data collator
data_collator = collator_module.DataCollatorSimpleSeamless(
    processor="facebook/hf-seamless-m4t-medium",
    max_audio_length_sec=8.0,
    max_text_length=256
)

# Prepare your data with translation and language pair information
your_data = [
    {
        'raw_audio': np.random.randn(16000 * 5),  # 5 seconds at 16kHz
        'raw_text': "Your subtitle text here",
        'is_translation': 1,       # 1 for translated content, 0 for original
        'language_pair_id': 5,     # 0-20 for specific language pairs
    }
]

# Process and run inference 
batch = data_collator(your_data)
model.eval()
with torch.no_grad():
    outputs = model(**batch)
    tte_prediction = outputs.logits.item()
    
print(f"Predicted Time To Edit (TTE): {tte_prediction:.2f} seconds")
```

## Model Details

- **Base Model**: SeamlessM4T (facebook/hf-seamless-m4t-medium)
- **Audio Encoder**: Frozen SeamlessM4T speech encoder  
- **Text Encoder**: Frozen SeamlessM4T text encoder
- **Hidden Size**: 1024
- **Translation Embedding**: 32 dimensions
- **Language Pair Embedding**: 64 dimensions  
- **Number of Language Pairs**: 21 (plus "other")
- **Audio Input**: 16kHz
- **Translation Input**: Binary flag (0/1)
- **Language Pair Input**: Categorical ID (0-20)
- **Output**: Single regression value (TTE in seconds)
- **Task**: Subtitle editing time prediction

## Supported Language Pairs

The model supports 21 specific translation pairs between 5 languages:

**Languages**: English (EN), French (FR), Spanish (ES), Italian (IT), German (DE)

**Translation Pairs**: All combinations between the 5 languages create various directional pairs (e.g., EN→FR, FR→EN, ES→IT, DE→ES, etc.). The model uses language pair IDs (0-20) to identify specific translation directions, with ID 21 reserved for "other" pairs.

## Data Format

Your input data should be a list of dictionaries with:
- `raw_audio`: NumPy array of audio samples (16kHz sampling rate)
- `raw_text`: String of subtitle text  
- `is_translation`: Binary flag (1 for translated, 0 for original content)
- `language_pair_id`: Integer ID (0-20) for specific language pair
- `labels`: Target TTE values in seconds (optional, for training)

Example:
```python
data = [
    {
        'raw_audio': audio_samples,  # shape: (num_samples,) at 16kHz
        'raw_text': "Subtitle text content",
        'is_translation': 1,     # 1 = translated, 0 = original
        'language_pair_id': 5,   # 0-20 for language pairs
        'labels': 2.5  # optional TTE target value in seconds
    }
]
```

## Performance Metrics

- **Best Eval RMSE**: 33.34

## Training Details

- **Base Model**: facebook/hf-seamless-m4t-medium
- **Model Type**: seamless_lang_pairs
- **Epochs**: 10
- **Batch Size (Train)**: 32
- **Batch Size (Eval)**: 64
- **Learning Rate**: 1.2e-4
- **LR Scheduler**: cosine_with_restarts
- **Warmup Ratio**: 0.05
- **Weight Decay**: 0.001
- **Optimizer**: AdamW (torch)
- **Max Grad Norm**: 1.0
- **FP16**: True
- **Early Stopping Patience**: 5
- **Audio Max Length**: 8.0 seconds
- **Text Max Length**: 256 tokens
- **Sample Rate**: 16kHz
- **Translation Feature**: Binary flag (0/1)
- **Language Pairs**: 21 pairs + other
- **Language Pair Embedding**: 64 dimensions
- **Normalization**: None (raw values)
- **Dataset Split**: 80/20 train/test
- **Random Seed**: 42
- **Metric**: RMSE (lower is better)

## Training Configuration

The model was trained with the following specifications:

- **Dataset**: Multimodal audio-subtitle pairs with translation and language pair annotations (5 languages: EN, FR, ES, IT, DE with 21 pairs)
- **Train/Test Split**: 80/20 with random seed 42
- **Audio Processing**: 16kHz sampling, max 8.0 seconds, no offset
- **Text Processing**: Max 256 tokens
- **Translation Feature**: Binary flag indicating original vs translated content
- **Language Pairs**: 21 translation pairs from 5 languages (EN, FR, ES, IT, DE) plus "other" category
- **Normalization**: None (raw TTE values in seconds)
- **Caching**: Audio segments cached and compressed for efficiency

## Usage Notes

- This is the **most advanced** variant with both translation and language pair features
- For simpler models, see `seamless-basic` (audio+text only) or `seamless-translation` (with translation flag)
- Model expects 16kHz audio input (automatically resampled by data collator)
- Both translation flag and language pair ID significantly impact predictions
- Language pair embeddings capture language-specific temporal patterns
- No feature normalization applied - outputs raw TTE predictions in seconds
- Optimized for fine-grained subtitle editing time estimation tasks

## Limitations

- Requires both translation and language pair annotations in training data
- Language pair embeddings are dataset-specific (top 21 pairs from training)
- Designed for TTE prediction, not general audio-text matching
- Performance may vary on out-of-domain content and unseen language pairs
- Requires specific data preprocessing (use included data collator)

## Related Models

- **[seamless-basic](https://huggingface.co/videoloc/seamless-basic)**: Basic audio+text model without translation or language features
- **[seamless-translation](https://huggingface.co/videoloc/seamless-translation)**: Includes translation awareness but no language pair embeddings
- **[seamless-crossattention](https://huggingface.co/videoloc/seamless-crossattention)**: Advanced cross-modal attention mechanisms for sophisticated audio-text interactions