piper-swift is a pure Swift + Metal runtime for the Piper VITS TTS model (macOS/iOS).
This project is a ground-up rewrite of the original MIT-licensed Piper project by Michael Hansen (aka synesthesiam) — see rhasspy/piper.
- No non-Swift runtime dependencies: the library does not depend on ONNX Runtime, Python, C++, CoreML, etc.
- GPU-first execution: keep tensors GPU-resident and execute operators via Metal compute kernels.
- Embeddable: integrate directly into macOS/iOS apps via SwiftPM (no separate native runtime to ship).
- License: MIT (see
LICENSE.md). - Attribution: this is a rewrite of the original MIT-licensed Piper project:
rhasspy/piper. - No GPL contamination: no part of this library was developed by referencing the GPL fork that the original Piper repo points to (the “piper1-gpl” successor). This codebase is fully MIT-licensed and will remain so.
swift build -c releaseswift run -c release piper-swift -- --voice en_GB-northern_english_male-medium --text "Hello from piper-swift"- Native Swift API + Metal acceleration with no external runtime
- Simple shipping story: one SwiftPM package; no Python toolchain; no
onnxruntimedylibs - Good scaling for longer synthesis workloads on Apple Silicon GPUs
All numbers below are wall-clock ms_mean from the included benchmarks using bench/fixtures/test_summary.json (base 14 phonemes, repeated by the scale factor). Swift runs were with PIPER_CPU_I64=1 and PIPER_METAL_BATCH=1.
| factor | phonemes | Swift/Metal (ms_mean) | ORT CPU (ms_mean) | ORT CoreML (ms_mean)* |
|---|---|---|---|---|
| 1 | 14 | 147.39 | 11.57 | 46.70 |
| 2 | 28 | 153.04 | 22.41 | 84.30 |
| 4 | 56 | 167.77 | 31.26 | 125.56 |
| 8 | 112 | 190.16 | 65.52 | 162.85 |
Scaling takeaway: Swift/Metal has higher fixed overhead at tiny inputs, but it grows more slowly as input length increases (the gap shrinks as phoneme count grows).
Derived from factor=1 → factor=8 (Δphonemes = +98):
| backend | ms_mean @1 | ms_mean @8 | growth (×) | Δms | Δms / Δphoneme |
|---|---|---|---|---|---|
| Swift/Metal | 147.39 | 190.16 | 1.29× | +42.77 | 0.44 |
| ORT CPU | 11.57 | 65.52 | 5.66× | +53.95 | 0.55 |
| ORT CoreML* | 46.70 | 162.85 | 3.49× | +116.15 | 1.19 |
* CoreMLExecutionProvider notes: ORT CoreML prints warnings about zero-length shape tensors and may segfault when run across multiple test vectors. The numbers above were collected with --max-tests 1.
- Swift scale benchmark:
PIPER_BENCH_GPU_TIMING=1 PIPER_CPU_I64=1 PIPER_METAL_BATCH=1 \
swift run -c release piper-swift --scale-bench --bench-summary bench/fixtures/test_summary.json --scale-factors 1,2,4,8 --max-tests 1- ONNX Runtime benchmark (optional; not required for the library):
python3 bench/benchmark_onnxruntime.py --summary bench/fixtures/test_summary.json --provider cpu --scale-factors 1,2,4,8 --max-tests 1
python3 bench/benchmark_onnxruntime.py --summary bench/fixtures/test_summary.json --provider coreml --scale-factors 1,2,4,8 --max-tests 1- Run benchmarks from clean clones:
chmod +x bench/run_from_clone.sh
./bench/run_from_clone.sh
