README.md

metadata

title: Diamond — Speech Restoration
emoji: 💎
colorFrom: blue
colorTo: indigo
sdk: gradio
sdk_version: 6.20.0
app_file: app.py
pinned: false
license: apache-2.0
short_description: Restore degraded speech to studio 44.1 kHz
models:
  - nineninesix/diamond-1.0

💎 Diamond — Speech Restoration

Turn degraded speech (podcast, phone, streamed, over-compressed) into near-studio 44.1 kHz audio.

Diamond is an RQ-Transformer: degraded speech → log-mel → Transformer encoder → time-transformer (causal self- and cross-attention) → depth-transformer (a small local autoregression over the 9 DAC RVQ codebooks per frame) → frozen DAC decoder → restored 44.1 kHz waveform. Decoding is chunked greedy with a repetition penalty to suppress autoregressive looping.

Runs on ZeroGPU — the model lives on CPU between requests and is moved to the GPU only for the duration of a restoration call.

• Model: nineninesix/diamond-1.0
• Training code: diamond-train
• Inference code: diamond-inference
• License: Apache-2.0