---
title: FocalCodec Quantization Lab
emoji: ⚡
colorFrom: blue
colorTo: purple
sdk: gradio
sdk_version: 6.30.0
app_file: app.py
short_description: FocalCodec weights at int4–int32 vs fp4–fp32
python_version: "3.12"
startup_duration_timeout: 1h
---

# ⚡ FocalCodec 12.5 Hz — weight quantization lab

Interactive demo of what happens to [`lucadellalib/focalcodec_12_5hz`](https://huggingface.co/lucadellalib/focalcodec_12_5hz)
when its **145 M weights** are stored at different numeric precisions:

`int4` · `int8` · `int16` · `int32` · `fp4` · `fp8` · `fp16` · `fp32`

The token stream is untouched, so the codec still runs at **0.1625 kbps**
(12.5 tokens/s × 13 bits) at every level — only the size of the network changes
(72.6 MB at 4-bit down to 581.1 MB at fp32).

## Three views

| Tab | What it shows |
| --- | --- |
| **Compare** | One reconstructed clip + input/reconstruction spectrograms, side by side for all 8 FocalCodec levels |
| **Playground** | Encode → decode a single clip at one chosen codec + weight format, with its metrics row |
| **Benchmark** | SI-SDR, STOI, LSD, stored model size, bitrate and real-time factor for all 18 levels — 8 FocalCodec + 8 LSCodec + 2 native-int8 rows — plus a summary chart |

## Method

Post-training **fake quantization** of every floating-point tensor: per-output-channel
absmax scale, round-to-nearest in the target format, dequantize back to fp32, then run
the untouched codec forward pass. `fp4` is IEEE-style E2M1, `fp8` is E4M3, `fp16`
and `fp32` are exact casts. Nothing is fine-tuned or calibrated on data.

The `int8-native` and `lscodec-int8-native` rows skip the dequantize step entirely:
the same int8 weights run through a real **W8A8 dynamic** runtime (`int8nn.py`,
shipped with the packed checkpoints) — weights stay int8 in memory, each GEMM
quantizes its activations per row and computes `int8 × int8 → int32` via
`torch._int_mm`. Their STOI/LSD track the plain `int8` rows.

The same sweep runs on a second codec, [LSCodec-25 Hz](https://github.com/X-LANCE/LSCodec-Inference),
whose rows are prefixed `lscodec-`: its encoder, vocoder and codebook are quantized
identically, while the WavLM-Large features it self-prompts with are not. It is a
different model at a different bitrate (0.25 kbps vs 0.1625 kbps), so read it as a
floor, not a control.

⚠️ **On SI-SDR:** it is deeply negative at every level, including `fp32`. That is
expected — at 0.16 kbps FocalCodec does not preserve waveform phase, and the
[FocalCodec paper](https://arxiv.org/abs/2502.04465) reports SI-SNR ≈ −37 dB for the
untouched model. Use **STOI** (intelligibility, higher is better) and **LSD**
(log-spectral distance, lower is better) to compare levels.

## Credits

Model, method and example audio by
[Luca Della Libera](https://huggingface.co/lucadellalib) et al. — Apache-2.0.
[FocalCodec: Low-Bitrate Speech Coding via Focal Modulation Networks](https://arxiv.org/abs/2502.04465), NeurIPS 2025.

The packed `int8` / `fp8` checkpoints produced by this lab live in
[`ALWOAL/focalcodec_12_5hz_quantized`](https://huggingface.co/ALWOAL/focalcodec_12_5hz_quantized),
and the packed LSCodec int8 files in [`ALWOAL/lscodec_25hz_int8`](https://huggingface.co/ALWOAL/lscodec_25hz_int8) —
both ship `int8nn.py` for native W8A8 inference.
The baseline codec is [LSCodec-Inference](https://github.com/X-LANCE/LSCodec-Inference) by
Yiwei Guo et al. — MIT.