Upload model

README.md

@@ -0,0 +1,199 @@
+---
+library_name: transformers
+tags: []
+---
+
+# Model Card for Model ID
+
+<!-- Provide a quick summary of what the model is/does. -->
+
+
+
+## Model Details
+
+### Model Description
+
+<!-- Provide a longer summary of what this model is. -->
+
+This is the model card of a 🤗 transformers model that has been pushed on the Hub. This model card has been automatically generated.
+
+- **Developed by:** [More Information Needed]
+- **Funded by [optional]:** [More Information Needed]
+- **Shared by [optional]:** [More Information Needed]
+- **Model type:** [More Information Needed]
+- **Language(s) (NLP):** [More Information Needed]
+- **License:** [More Information Needed]
+- **Finetuned from model [optional]:** [More Information Needed]
+
+### Model Sources [optional]
+
+<!-- Provide the basic links for the model. -->
+
+- **Repository:** [More Information Needed]
+- **Paper [optional]:** [More Information Needed]
+- **Demo [optional]:** [More Information Needed]
+
+## Uses
+
+<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->
+
+### Direct Use
+
+<!-- This section is for the model use without fine-tuning or plugging into a larger ecosystem/app. -->
+
+[More Information Needed]
+
+### Downstream Use [optional]
+
+<!-- This section is for the model use when fine-tuned for a task, or when plugged into a larger ecosystem/app -->
+
+[More Information Needed]
+
+### Out-of-Scope Use
+
+<!-- This section addresses misuse, malicious use, and uses that the model will not work well for. -->
+
+[More Information Needed]
+
+## Bias, Risks, and Limitations
+
+<!-- This section is meant to convey both technical and sociotechnical limitations. -->
+
+[More Information Needed]
+
+### Recommendations
+
+<!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->
+
+Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.
+
+## How to Get Started with the Model
+
+Use the code below to get started with the model.
+
+[More Information Needed]
+
+## Training Details
+
+### Training Data
+
+<!-- This should link to a Dataset Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->
+
+[More Information Needed]
+
+### Training Procedure
+
+<!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the training procedure. -->
+
+#### Preprocessing [optional]
+
+[More Information Needed]
+
+
+#### Training Hyperparameters
+
+- **Training regime:** [More Information Needed] <!--fp32, fp16 mixed precision, bf16 mixed precision, bf16 non-mixed precision, fp16 non-mixed precision, fp8 mixed precision -->
+
+#### Speeds, Sizes, Times [optional]
+
+<!-- This section provides information about throughput, start/end time, checkpoint size if relevant, etc. -->
+
+[More Information Needed]
+
+## Evaluation
+
+<!-- This section describes the evaluation protocols and provides the results. -->
+
+### Testing Data, Factors & Metrics
+
+#### Testing Data
+
+<!-- This should link to a Dataset Card if possible. -->
+
+[More Information Needed]
+
+#### Factors
+
+<!-- These are the things the evaluation is disaggregating by, e.g., subpopulations or domains. -->
+
+[More Information Needed]
+
+#### Metrics
+
+<!-- These are the evaluation metrics being used, ideally with a description of why. -->
+
+[More Information Needed]
+
+### Results
+
+[More Information Needed]
+
+#### Summary
+
+
+
+## Model Examination [optional]
+
+<!-- Relevant interpretability work for the model goes here -->
+
+[More Information Needed]
+
+## Environmental Impact
+
+<!-- Total emissions (in grams of CO2eq) and additional considerations, such as electricity usage, go here. Edit the suggested text below accordingly -->
+
+Carbon emissions can be estimated using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute) presented in [Lacoste et al. (2019)](https://arxiv.org/abs/1910.09700).
+
+- **Hardware Type:** [More Information Needed]
+- **Hours used:** [More Information Needed]
+- **Cloud Provider:** [More Information Needed]
+- **Compute Region:** [More Information Needed]
+- **Carbon Emitted:** [More Information Needed]
+
+## Technical Specifications [optional]
+
+### Model Architecture and Objective
+
+[More Information Needed]
+
+### Compute Infrastructure
+
+[More Information Needed]
+
+#### Hardware
+
+[More Information Needed]
+
+#### Software
+
+[More Information Needed]
+
+## Citation [optional]
+
+<!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->
+
+**BibTeX:**
+
+[More Information Needed]
+
+**APA:**
+
+[More Information Needed]
+
+## Glossary [optional]
+
+<!-- If relevant, include terms and calculations in this section that can help readers understand the model or model card. -->
+
+[More Information Needed]
+
+## More Information [optional]
+
+[More Information Needed]
+
+## Model Card Authors [optional]
+
+[More Information Needed]
+
+## Model Card Contact
+
+[More Information Needed]

config.json

@@ -0,0 +1,17 @@
+{
+  "architectures": [
+    "ASRRouterModel"
+  ],
+  "auto_map": {
+    "AutoConfig": "router_model.ASRRouterConfig",
+    "AutoModel": "router_model.ASRRouterModel"
+  },
+  "dropout": 0.1,
+  "dtype": "float32",
+  "hidden_dim": 128,
+  "input_dim": 384,
+  "intermediate_dim": 64,
+  "model_type": "asr_router",
+  "num_models": 3,
+  "transformers_version": "4.57.3"
+}

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:23e194a9918f01d9c552d694f21dedd5d2b9566713644bc313152dac34127be2
+size 233284

router_model.py

@@ -0,0 +1,202 @@
+"""
+Router Model Architecture for Smart ASR Routing.
+
+Regression-based approach: predicts WER for each backend model.
+"""
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from dataclasses import dataclass
+from typing import Optional, Dict
+
+from transformers import PreTrainedModel, PretrainedConfig, WhisperModel, WhisperFeatureExtractor
+from transformers.modeling_outputs import ModelOutput
+
+
+class AttentionPooling(nn.Module):
+    """Learnable attention pooling for variable-length sequences."""
+
+    def __init__(self, input_dim: int):
+        super().__init__()
+        self.attention = nn.Sequential(
+            nn.Linear(input_dim, 1),
+            nn.Tanh()
+        )
+
+    def forward(self, x: torch.Tensor, mask: torch.Tensor = None) -> torch.Tensor:
+        """
+        Args:
+            x: [Batch, Time, Dim]
+            mask: [Batch, Time] (1 for valid, 0 for pad)
+        Returns:
+            pooled: [Batch, Dim]
+        """
+        scores = self.attention(x)  # [Batch, Time, 1]
+
+        if mask is not None:
+            scores = scores.masked_fill(mask.unsqueeze(-1) == 0, -1e9)
+
+        weights = F.softmax(scores, dim=1)  # [Batch, Time, 1]
+        return torch.sum(x * weights, dim=1)  # [Batch, Dim]
+
+
+class ASRRouterConfig(PretrainedConfig):
+    """Configuration for ASRRouter model."""
+    model_type = "asr_router"
+
+    def __init__(
+        self,
+        input_dim: int = 384,  # whisper-tiny encoder dim
+        hidden_dim: int = 128,
+        intermediate_dim: int = 64,
+        dropout: float = 0.1,  # Lower dropout for regression
+        num_models: int = 3,   # Number of backends to predict scores for
+        **kwargs
+    ):
+        super().__init__(**kwargs)
+        self.input_dim = input_dim
+        self.hidden_dim = hidden_dim
+        self.intermediate_dim = intermediate_dim
+        self.dropout = dropout
+        self.num_models = num_models
+
+
+@dataclass
+class RouterOutput(ModelOutput):
+    loss: Optional[torch.FloatTensor] = None
+    pred_wers: torch.FloatTensor = None  # Predicted WER for each model
+
+
+class ASRRouterModel(PreTrainedModel):
+    """
+    Regression Router.
+    Input: 384-dimensional Whisper encoder embeddings
+    Output: Estimated WER (0.0+, unbounded) for each backend model.
+    Uses Softplus activation to ensure non-negative outputs while allowing WER > 1.0.
+    """
+    config_class = ASRRouterConfig
+
+    MODEL_ID_MAP = {0: "kyutai", 1: "granite", 2: "tiny_audio"}
+
+    def __init__(self, config: ASRRouterConfig):
+        super().__init__(config)
+
+        self.network = nn.Sequential(
+            nn.Linear(config.input_dim, config.hidden_dim),
+            nn.GELU(),
+            nn.LayerNorm(config.hidden_dim),  # Better for batch_size=1
+            nn.Dropout(config.dropout),
+
+            nn.Linear(config.hidden_dim, config.intermediate_dim),
+            nn.GELU(),
+            nn.LayerNorm(config.intermediate_dim),
+
+            nn.Linear(config.intermediate_dim, config.num_models)
+        )
+
+        self.post_init()
+
+    def forward(
+        self,
+        embeddings: torch.Tensor,
+        labels: Optional[torch.Tensor] = None,  # Actual WERs from ground truth
+    ) -> RouterOutput:
+
+        # Softplus for unbounded positive WER (WER can exceed 1.0)
+        pred_wers = F.softplus(self.network(embeddings))
+
+        loss = None
+        if labels is not None:
+            loss = F.mse_loss(pred_wers, labels)
+
+        return RouterOutput(loss=loss, pred_wers=pred_wers)
+
+    def predict_proba(self, embeddings: torch.Tensor) -> torch.Tensor:
+        """Get predicted WERs for each model."""
+        with torch.no_grad():
+            return F.softplus(self.network(embeddings))
+
+
+class RouterWithFeatureExtractor:
+    """
+    Production-ready router with attention pooling and memory optimizations.
+    """
+    def __init__(self, router: ASRRouterModel, device: str = "cpu"):
+        self.device = device
+        self.router = router.to(device)
+        self.router.eval()
+
+        # Attention pooling for variable-length sequences
+        self.attention_pooling = AttentionPooling(input_dim=384).to(device)
+        self.attention_pooling.eval()
+
+        # Memory Optimization: Load full model, extract encoder, delete rest
+        print("Loading Whisper Encoder...")
+        full_whisper = WhisperModel.from_pretrained("openai/whisper-tiny")
+        self.whisper_encoder = full_whisper.encoder.to(device)
+        self.whisper_encoder.eval()
+
+        del full_whisper.decoder
+        del full_whisper
+        torch.cuda.empty_cache() if torch.cuda.is_available() else None
+
+        self.feature_extractor = WhisperFeatureExtractor.from_pretrained("openai/whisper-tiny")
+
+    def extract_features(self, waveform: torch.Tensor) -> torch.Tensor:
+        """Extract embeddings using Attention Pooling for variable lengths."""
+        if waveform.dim() == 1:
+            waveform = waveform.unsqueeze(0)
+
+        # Convert batch tensor to list of 1D numpy arrays (required by WhisperFeatureExtractor)
+        audio_np = [w.cpu().numpy() for w in waveform]
+
+        inputs = self.feature_extractor(
+            audio_np,
+            sampling_rate=16000,
+            return_tensors="pt",
+            return_attention_mask=True
+        )
+
+        input_features = inputs.input_features.to(self.device)
+        attention_mask = inputs.attention_mask.to(self.device)
+
+        with torch.no_grad():
+            last_hidden_state = self.whisper_encoder(input_features).last_hidden_state
+
+            # Resize mask to match encoder output temporal dimension
+            mask_resized = F.interpolate(
+                attention_mask.unsqueeze(1).float(),
+                size=last_hidden_state.shape[1],
+                mode='nearest'
+            ).squeeze(1)
+
+            # Attention Pooling
+            return self.attention_pooling(last_hidden_state, mask_resized)
+
+    def predict(self, waveform: torch.Tensor) -> Dict:
+        """Select the model with the lowest predicted WER."""
+        embeddings = self.extract_features(waveform)
+
+        with torch.no_grad():
+            output = self.router(embeddings)
+            pred_wers = output.pred_wers[0].cpu().numpy()
+
+        scores = {
+            "kyutai": float(pred_wers[0]),
+            "granite": float(pred_wers[1]),
+            "tiny_audio": float(pred_wers[2])
+        }
+
+        best_model = min(scores.items(), key=lambda x: x[1])
+
+        return {
+            "selected_model": best_model[0],
+            "predicted_wers": scores,
+            "confidence": max(0.0, 1.0 - best_model[1])  # Clamp since WER can exceed 1.0
+        }
+
+
+# Register for auto classes
+ASRRouterConfig.register_for_auto_class()
+ASRRouterModel.register_for_auto_class("AutoModel")