Update custom model files, README, and requirements

README.md

@@ -1,84 +1,71 @@
 ---
-library_name: transformers
+license: mit
+language:
+- en
+datasets:
+- speechbrain/LoquaciousSet
+base_model:
+- openai/whisper-large-v3-turbo
+- HuggingFaceTB/SmolLM3-3B
+pipeline_tag: automatic-speech-recognition
 tags:
-- generated_from_trainer
-model-index:
-- name: tiny-audio
-  results: []
+- asr
+- speech-recognition
+- audio
+- smollm
+- whisper
+- mlp
 ---
 
-<!-- This model card has been generated automatically according to the information the Trainer had access to. You
-should probably proofread and complete it, then remove this comment. -->
-
-# tiny-audio
-
-This model is a fine-tuned version of [](https://huggingface.co/) on an unknown dataset.
-It achieves the following results on the evaluation set:
-- Loss: 0.3030
-
-## Model description
-
-More information needed
-
-## Intended uses & limitations
-
-More information needed
-
-## Training and evaluation data
-
-More information needed
-
-## Training procedure
-
-### Training hyperparameters
-
-The following hyperparameters were used during training:
-- learning_rate: 0.0002
-- train_batch_size: 20
-- eval_batch_size: 20
-- seed: 42
-- gradient_accumulation_steps: 2
-- total_train_batch_size: 40
-- optimizer: Use OptimizerNames.ADAMW_TORCH_FUSED with betas=(0.9,0.95) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
-- lr_scheduler_type: cosine
-- lr_scheduler_warmup_steps: 500
-- num_epochs: 1
-
-### Training results
-
-| Training Loss | Epoch  | Step  | Validation Loss |
-|:-------------:|:------:|:-----:|:---------------:|
-| 0.5056        | 0.0372 | 1000  | 0.4229          |
-| 0.4725        | 0.0745 | 2000  | 0.3813          |
-| 0.4602        | 0.1117 | 3000  | 0.3660          |
-| 0.4293        | 0.1490 | 4000  | 0.3542          |
-| 0.4113        | 0.1862 | 5000  | 0.3468          |
-| 0.4153        | 0.2234 | 6000  | 0.3404          |
-| 0.3996        | 0.2607 | 7000  | 0.3337          |
-| 0.4089        | 0.2979 | 8000  | 0.3294          |
-| 0.3743        | 0.3352 | 9000  | 0.3250          |
-| 0.4082        | 0.3724 | 10000 | 0.3214          |
-| 0.3679        | 0.4096 | 11000 | 0.3193          |
-| 0.3792        | 0.4469 | 12000 | 0.3185          |
-| 0.3723        | 0.4841 | 13000 | 0.3140          |
-| 0.3696        | 0.5214 | 14000 | 0.3109          |
-| 0.3920        | 0.5586 | 15000 | 0.3104          |
-| 0.3827        | 0.5958 | 16000 | 0.3070          |
-| 0.3593        | 0.6331 | 17000 | 0.3054          |
-| 0.3832        | 0.6703 | 18000 | 0.3056          |
-| 0.3772        | 0.7076 | 19000 | 0.3038          |
-| 0.3887        | 0.7448 | 20000 | 0.3035          |
-| 0.3666        | 0.7821 | 21000 | 0.3034          |
-| 0.3690        | 0.8193 | 22000 | 0.3033          |
-| 0.3568        | 0.8565 | 23000 | 0.3031          |
-| 0.3809        | 0.8938 | 24000 | 0.3029          |
-| 0.3736        | 0.9310 | 25000 | 0.3032          |
-| 0.3553        | 0.9683 | 26000 | 0.3030          |
-
-
-### Framework versions
-
-- Transformers 5.0.0.dev0
-- Pytorch 2.8.0+cu128
-- Datasets 3.6.0
-- Tokenizers 0.22.1
+# Tiny Audio
+
+A speech recognition model trained in 24 hours on a single GPU for ~$12. Built with the [Tiny Audio](https://github.com/alexkroman/tiny-audio) codebase—a minimal, hackable framework for training ASR models.
+
+## Architecture
+
+```
+Audio (16kHz) → Whisper Encoder (frozen) → MLP Projector (trained) → SmolLM3-3B (frozen) → Text
+```
+
+**MLP Projector:**
+- Convolutional downsampling: 4x sequence compression via two stride-2 conv layers
+- Linear (1280 → 2048) → GELU → Linear (2048 → 2048)
+- Output normalization: RMSNorm
+
+## Training Details
+
+| | |
+|---|---|
+| **Dataset** | LoquaciousSet (25,000 hours) |
+| **Hardware** | Single NVIDIA A40 40GB |
+| **Training Time** | ~24 hours |
+| **Cost** | ~$12 |
+| **Trainable Parameters** | ~12M (projector only) |
+
+## Performance
+
+**Word Error Rate (WER): 12.14%** on LoquaciousSet test set.
+
+
+## Usage
+
+```python
+from transformers import pipeline
+
+pipe = pipeline("automatic-speech-recognition", model="mazesmazes/tiny-audio", trust_remote_code=True)
+
+result = pipe("path/to/audio.wav")
+print(result["text"])
+```
+
+## Limitations
+
+- English only
+- Optimized for 16kHz audio; other sample rates are resampled automatically
+- Performance may degrade on heavily accented speech, noisy environments, or domain-specific jargon
+- Maximum audio length limited by context window
+
+## Learn More
+
+- **[Train your own model](https://github.com/alexkroman/tiny-audio)** — The full codebase with training scripts
+- **[Free 3.5-hour course](https://github.com/alexkroman/tiny-audio/blob/main/docs/course/0-course-overview.md)** — Build your own ASR system from scratch

asr_config.py

@@ -71,7 +71,7 @@ class ASRConfig(transformers.PretrainedConfig):
             "min_new_tokens": 0,
             "repetition_penalty": 1.0,
             "length_penalty": 1.0,
-            "no_repeat_ngram_size": 0,
+            "no_repeat_ngram_size": 3,  # Prevent repeating 3-grams like "so so so"
             "use_cache": True,
         }
 

asr_pipeline.py

@@ -486,6 +486,7 @@ class ASRPipeline(transformers.AutomaticSpeechRecognitionPipeline):
             return ""
 
         original_len = len(text.split())
+        original_text = text  # Keep for debug
 
         # 1. LOWERCASE
         text = text.lower()
@@ -505,8 +506,10 @@ class ASRPipeline(transformers.AutomaticSpeechRecognitionPipeline):
                 words = words[: idx + n]
                 text = " ".join(words)
                 print(
-                    f"[DEBUG] Truncated repetition: {original_len} -> {len(words)} words (n={n}, repeats={repeat_count})"
+                    f"[POSTPROCESS] Truncated repetition: {original_len} -> {len(words)} words (n={n}, repeats={repeat_count})"
                 )
+                print(f"[POSTPROCESS] Before: {original_text[:100]}...")
+                print(f"[POSTPROCESS] After: {text[:100]}...")
                 break
 
         # 3. COMBINE ACRONYMS