Add model card with val loss table and usage example

README.md

@@ -0,0 +1,106 @@
+---
+base_model: openbmb/VoxCPM2
+language:
+- dag
+language_bcp47:
+- dag-gh
+tags:
+- text-to-speech
+- full-finetune
+- dagbani
+- farmerline
+- african-languages
+license: apache-2.0
+---
+
+# VoxCPM2 Dagbani — Full SFT
+
+Full fine-tune of [VoxCPM2](https://huggingface.co/openbmb/VoxCPM2) on Dagbani  speech data. Trained by [FarmerlineML](https://huggingface.co/FarmerlineML) for the darli AI agricultural voice assistant.
+
+## Training Details
+
+| Parameter | Value |
+|---|---|
+| Base model | openbmb/VoxCPM2 (2B) |
+| Method | Full SFT (all parameters) |
+| Learning rate | 1e-5 |
+| Batch size | 1 (grad accum 16, effective batch 16) |
+| Sample rate | 16kHz (AudioVAE encoder input) |
+| Final step | 3062 |
+
+**Validation Loss:**
+
+| Step | loss/total | loss/diff | loss/stop |
+|------|-----------|-----------|----------|
+| 0 | 1.107430 | 0.857099 | 0.166888 |
+| 500 | 0.868330 | 0.819919 | 0.032275 |
+| 1000 | 0.864477 | 0.812996 | 0.034321 |
+| 1500 | 0.855272 | 0.806172 | 0.032734 |
+| 2000 | 0.870591 | 0.798133 | 0.048305 |
+| 2500 | 0.865124 | 0.800481 | 0.043095 |
+| 3000 | 0.851482 | 0.798783 | 0.035133 |
+
+**Datasets:**
+- [ghananlpcommunity/navigation-corpus-dagbani-speech](https://huggingface.co/datasets/ghananlpcommunity/navigation-corpus-dagbani-speech)
+- [FarmerlineML/DagbaniTTS2025_dataset](https://huggingface.co/datasets/FarmerlineML/DagbaniTTS2025_dataset)
+
+## Usage
+
+```python
+from voxcpm import VoxCPM
+import soundfile as sf
+import numpy as np
+
+model = VoxCPM.from_pretrained(
+    "FarmerlineML/voxcpm2-dagbani-sft",
+    load_denoiser=False,
+)
+
+def trim_audio(wav, sr, silence_thresh=0.01, max_silence_secs=2.0):
+    abs_wav = np.abs(wav)
+    window  = int(0.05 * sr)
+    n_wins  = len(abs_wav) // window
+    max_sil = int(max_silence_secs / 0.05)
+    silence_count, cut_sample = 0, len(wav)
+    for w in range(n_wins):
+        chunk = abs_wav[w * window:(w + 1) * window]
+        if chunk.max() < silence_thresh:
+            silence_count += 1
+            if silence_count >= max_sil:
+                cut_sample = (w - max_sil + 1) * window
+                break
+        else:
+            silence_count = 0
+    return wav[:min(cut_sample + int(0.1 * sr), len(wav))]
+
+wav = model.generate(
+    text="a nyɛla Dagbanli lɔri yubu daluu",
+    reference_wav_path="your_dagbani_speaker.wav",
+    cfg_value=2.0,
+    inference_timesteps=15,
+    retry_badcase=False,
+    max_len=max(50, len(text) * 4),
+)
+wav = trim_audio(wav, 48000)
+sf.write("output.wav", wav, 48000)
+```
+
+## Repo Structure
+
+```
+├── model.safetensors        # Model weights (~9.2GB)
+├── audiovae.pth             # AudioVAE decoder
+├── config.json              # Model architecture config
+├── tokenizer.json           # Tokenizer
+├── training/
+│   ├── train.log            # Full training log
+│   ├── val_loss_summary.txt # Validation losses per checkpoint
+│   └── training_state.json  # Final training state
+└── tensorboard/             # TensorBoard event files
+```
+
+## Notes
+
+- Reference audio is required at inference for voice identity anchoring
+- Use `max_len=max(50, len(text) * 4)` to prevent hallucination after sentence end
+- A post-generation 2-second silence trim is strongly recommended