Text-to-Speech
Safetensors
English
steertts
steer-tts
tts
speech-synthesis
english
controllable-tts
steerable-tts
prosody-control
expressive-tts
style-control
inline-markup
word-emphasis
zero-shot-voice-cloning
voice-cloning
interpretable
non-autoregressive
fastpitch
mamba
mamba-3
state-space-model
hybrid-attention
flow-matching
vocos
small-model
quantized
int8
on-device
streaming
non-commercial
Eval Results (legacy)
| import numpy as np | |
| import pytest | |
| from steertts import SR | |
| TEXT = "The meeting has been moved to Thursday afternoon at three." | |
| def tts8(model_dir): | |
| from pathlib import Path | |
| if not (Path(model_dir) / "int8" / "model.safetensors").exists(): | |
| pytest.skip("no int8/ weights in the model directory") | |
| from steertts import SteerTTS | |
| from conftest import DEV | |
| return SteerTTS(model_dir, device=DEV, flow=False, precision="int8") | |
| def test_int8_render(tts8): | |
| a = tts8(TEXT) | |
| assert a.dtype == np.float32 and 1.5 < len(a) / SR < 8 and np.isfinite(a).all() and np.abs(a).max() > 0.01 | |
| assert np.array_equal(a, tts8(TEXT)) | |
| def test_int8_close_to_fp32(tts, tts8): | |
| assert abs(len(tts8(TEXT)) / len(tts(TEXT)) - 1) < 0.05 | |
| sd8, sd = tts8.model.state_dict(), tts.model.state_dict() | |
| for k, w in sd.items(): | |
| if w.is_floating_point() and w.numel(): | |
| assert (sd8[k] - w).abs().max() <= w.abs().max() * 4e-3 + 1e-6, k | |
| def test_int8_readback_direction(tts8): | |
| pytest.importorskip("parselmouth") | |
| from steertts.measure import measure_z | |
| v = tts8.default_voice | |
| z = {s: measure_z(tts8(TEXT, ctrl={"f0_med": s}), TEXT, tts8.voice_stats(v), tts8.pooled_sd)["f0_med"] for s in (-2, 2)} | |
| assert z[2] - z[-2] > 1.0, z | |
| def test_round_int8_matches_packed(model_dir): | |
| """The in-place rounding used for the external models equals the packed int8 files (same rule).""" | |
| import torch | |
| from steertts.quant import load_packed, round_int8_ | |
| from safetensors.torch import load_file | |
| sd = load_file(f"{model_dir}/model.safetensors") | |
| packed = load_packed(f"{model_dir}/int8/model.safetensors") | |
| m = torch.nn.Module() | |
| for k, w in sd.items(): | |
| m.register_parameter(k.replace(".", "_"), torch.nn.Parameter(w.clone(), requires_grad=False)) | |
| round_int8_(m) | |
| for k, w in sd.items(): | |
| assert torch.equal(getattr(m, k.replace(".", "_")), packed[k]), k | |