rlx-tsac

Native Rust TSAC for RLX plus optional wrapper around Fabrice Bellard's TSAC Linux binary — very low bitrate neural audio (~5–8 kb/s @ 44.1 kHz).

The native codec (vendored tsac-ng via FFI) runs on the host with RLX device tags (cpu, metal, cuda, vulkan, …). Reference binaries (original Bellard tsac + standalone tsac-ng) run in a small Docker image for linux/amd64 perf and parity on any host.

Quick start

just fetch-tsac
just tsac -- --in-wav speech.wav --out-wav /tmp/out.wav --device cpu

Perf bench (RLX host vs Docker refs)

RLX runs on the host; Docker only runs the original Bellard binary and standalone tsac-ng:

bash crates/rlx-tsac/docker/run.sh build

cargo run -p rlx-tsac --example perf_bench --release --features native-codec -- \
  --in-wav speech.wav --quality 9 --fast --device metal

Prints encode/decode timing for all three engines plus PCM correlation vs the Docker references.

Tolerances: RLX_TSAC_PARITY_MIN_CORR (default 0.92), RLX_TSAC_PARITY_MAX_MSE (default 0.0025).

Bellard parity (Linux x86_64 host, no Docker)

On Linux x86_64 you can compare native RLX against the Bellard binary directly:

just bench-tsac-parity
just bench-tsac-parity-example -- --in-wav speech.wav --quality 9 --fast

Library

use rlx_tsac::{TsacCodec, TsacBackendKind, TsacOptions, bench_perf, PerfOptions};
use rlx_runtime::Device;

let native = TsacCodec::open_native(".cache/tsac", TsacOptions {
    backend: TsacBackendKind::Native,
    device: Device::Metal,
    ..Default::default()
})?;

let report = bench_perf(".cache/tsac", "speech.wav", &PerfOptions::default())?;
report.print_summary(&PerfOptions::default());

Environment

Options (passed to tsac)

Tests

cargo test -p rlx-tsac --release

# Linux x86_64 Bellard parity (host, no Docker):
RLX_TSAC_PARITY=1 cargo test -p rlx-tsac --test bellard_parity --release --features fetch,native-codec -- --nocapture