| |
| """CuteTTS (AX650) command line demo. |
| |
| python3 demo.py --model-dir models --text "hello" --reference ref.wav --output out.wav |
| """ |
| import argparse |
| import os |
| import sys |
|
|
| import numpy as np |
|
|
| sys.path.insert(0, os.path.dirname(os.path.abspath(__file__))) |
|
|
| from cutetts_ax import CuteTTSAX |
| from cutetts_ax import audio as A |
|
|
|
|
| def main(): |
| ap = argparse.ArgumentParser() |
| ap.add_argument("--model-dir", default=os.environ.get("CUTETTS_MODEL_DIR", "models")) |
| ap.add_argument("--assets", default=os.path.join(os.path.dirname(os.path.abspath(__file__)), "assets")) |
| ap.add_argument("--engine", default="ax", choices=["ax", "onnx"]) |
| ap.add_argument("--text", default="The voice generated by this model sounds amazing!") |
| ap.add_argument("--reference", default="samples/reference.wav") |
| ap.add_argument("--output", default="output.wav") |
| ap.add_argument("--cfg", type=float, default=None) |
| ap.add_argument("--max-new", type=int, default=None) |
| ap.add_argument("--seed", type=int, default=42) |
| ap.add_argument("--ref-seconds", type=float, default=5.0) |
| ap.add_argument("--stream", action="store_true", help="true streaming: emit audio while generating") |
| ap.add_argument("--first-chunk", type=int, default=16, help="latents in the first streamed chunk") |
| ap.add_argument("--chunk", type=int, default=32, help="latents per streamed chunk afterwards") |
| ap.add_argument("--windowed", action="store_true", |
| help="exact windowed streaming (bit-identical to offline, higher latency)") |
| a = ap.parse_args() |
|
|
| tts = CuteTTSAX(a.model_dir, a.assets, engine=a.engine) |
| if a.stream: |
| import wave |
| with wave.open(a.output, "wb") as wf: |
| wf.setnchannels(1) |
| wf.setsampwidth(2) |
| wf.setframerate(tts.sr) |
| for chunk in tts.synthesize_stream(a.text, a.reference, max_new=a.max_new, |
| cfg_strength=a.cfg, seed=a.seed, |
| ref_max_seconds=a.ref_seconds, |
| first_chunk=a.first_chunk, chunk=a.chunk, |
| windowed=a.windowed): |
| pcm = (np.clip(chunk, -1.0, 1.0) * 32767.0).astype("<i2") |
| wf.writeframes(pcm.tobytes()) |
| tm = tts.last_timing |
| print(f"streaming: {tm['chunks']} chunks, TTFT {tm['ttft']:.2f}s, " |
| f"audio {tm['audio_seconds']:.2f}s in {tm['total']:.2f}s " |
| f"(RTF {tm['total']/max(tm['audio_seconds'],1e-6):.3f})") |
| print("saved", a.output) |
| return |
| wav = tts.synthesize(a.text, a.reference, max_new=a.max_new, cfg_strength=a.cfg, |
| seed=a.seed, ref_max_seconds=a.ref_seconds) |
| A.write_wav(a.output, wav, tts.sr) |
| tm = tts.last_timing |
| print(f"audio {tm['audio_seconds']:.2f}s in {tm['total']:.2f}s " |
| f"(RTF {tm['total']/max(tm['audio_seconds'],1e-6):.2f}), {tm['steps']} patches") |
| print(" vae_enc %.3f speaker %.3f prompt %.3f prefill %.3f ar %.3f " |
| "(dit %.3f, lm_dec %.3f) vae_dec %.3f" |
| % (tm["vae_enc"], tm["speaker_enc"], tm["prompt"], tm["prefill"], tm["ar_loop"], |
| tm["dit"], tm["lm_decode"], tm["vae_dec"])) |
| print("saved", a.output) |
|
|
|
|
| if __name__ == "__main__": |
| main() |
|
|
