Modelpublic

naazimsnh02/voiceown-base-gguf

sourceHugging Faceapache-2.0updated 5mo agoView on Hugging Face

1likes106downloads

Model Card

VoiceOwn — Stutter-Aware Intelligence

VoiceOwn is a specialized adaptation of Gemma 4 E2B, engineered to bridge communication gaps for individuals who stutter.

Unlike conventional ASR systems that transcribe every disfluency, this model performs intent-focused speech understanding—filtering repetitions, prolongations, and blocks to produce clean, intended language.

✨ Core Capabilities

—Stutter-Awareness Handles repetitions, prolongations, and speech blocks natively.

—Intent Extraction Identifies the speaker’s intended words rather than literal disfluent output.

—Multimodal Intelligence Uses Gemma 4’s audio encoder to interpret timing, tone, and structure of speech.

📦 Model Weights (GGUF)

File	Description
`voiceown-base-Q4_K_M.gguf`	Mobile-optimized
`voiceown-base-Q8_0.gguf`	Higher-fidelity evaluation build
`gemma-4-e2b-it.BF16-mmproj.gguf`	Required multimodal projector (core logic)

🧪 Training Insights

—Objective: Intent-accurate output from disfluent speech
—Dataset: naazimsnh02/voiceown-stutter-asr
—Samples: 2,850 real-world recordings
—Epochs: 2
—Training Loss: 1.1124
—Hardware: NVIDIA A100-SXM4
—Training Time: ~60 minutes

⚙️ Usage

Run with llama.cpp multimodal CLI:

bash

./llama-mtmd-cli \
  -m voiceown-base-Q4_K_M.gguf \
  --mmproj gemma-4-e2b-it.BF16-mmproj.gguf \
  --audio user_clip.wav \
  -p "Capture the speaker's intended words, ignoring any stutters."