| """Mach Boom: text to 24 kHz speech on a CPU. |
| |
| python tts.py "Your booking reference is AB-4471-Z." -o out.wav --voice Zoey |
| """ |
| import argparse, os, sys, time, wave |
| import numpy as np |
| HERE = os.path.dirname(os.path.abspath(__file__)) |
| sys.path.insert(0, HERE) |
| from mach import MachBoom |
|
|
|
|
| def pcm16(x): |
| return (np.clip(x, -1, 1) * 32767).round().astype("<i2").tobytes() |
|
|
|
|
| def write_wav(path, x, sr=24000): |
| with wave.open(path, "wb") as w: |
| w.setnchannels(1); w.setsampwidth(2); w.setframerate(sr); w.writeframes(pcm16(x)) |
|
|
|
|
| def main(): |
| ap = argparse.ArgumentParser(description=__doc__.split("\n")[0]) |
| ap.add_argument("text") |
| ap.add_argument("-o", "--out", default="out.wav") |
| ap.add_argument("--voice", default="Zoey") |
| ap.add_argument("--model", default=os.path.join(HERE, "mach-boom.bin")) |
| ap.add_argument("--tokenizer", default=os.path.join(HERE, "tokenizer.json")) |
| ap.add_argument("--voices", default=os.path.join(HERE, "voices")) |
| ap.add_argument("--threads", type=int, default=1, help="1 to 4: the audio decoder runs alongside the speech model") |
| ap.add_argument("--seed", type=int, default=0) |
| ap.add_argument("--chunk", type=int, default=None, help="decoder chunk in frames (default 16, or 8 above 2 threads)") |
| ap.add_argument("--no-frontend", action="store_true", help="skip the English text normalizer") |
| ap.add_argument("--stream", action="store_true", help="synthesize piece by piece and print the time to first audio") |
| ap.add_argument("--first", type=int, default=None, help="frames in the first piece when streaming (2: first audio " |
| "about 2x sooner, at about 50 dB from the default chunking)") |
| ap.add_argument("--pcm", action="store_true", help="stream raw 16-bit 24 kHz mono PCM to stdout, piece by piece") |
| a = ap.parse_args() |
| tts = MachBoom(a.model, voices_dir=a.voices, tokenizer=a.tokenizer) |
| args = dict(seed=a.seed, frontend=not a.no_frontend, threads=a.threads, chunk=a.chunk) |
| t0 = time.perf_counter() |
| if a.stream or a.pcm: |
| pieces, log = [], sys.stderr if a.pcm else sys.stdout |
| for x in tts.synth_stream(a.text, a.voice, first=a.first, **args): |
| if a.pcm: |
| sys.stdout.buffer.write(pcm16(x)); sys.stdout.buffer.flush() |
| if not pieces: |
| print(f"first audio after {(time.perf_counter() - t0) * 1e3:.0f} ms ({len(x) / 24:.0f} ms of audio)", file=log) |
| pieces.append(x) |
| wav = np.concatenate(pieces) |
| else: |
| wav = tts.synth(a.text, a.voice, **args) |
| dt = time.perf_counter() - t0 |
| write_wav(a.out, wav) |
| msg = f"{a.out}: {len(wav) / 24000:.2f} s of audio in {dt:.2f} s ({len(wav) / 24000 / dt:.1f}x real time, {tts.k.isa}" |
| print(msg + (" + amx)" if tts.amx else ")"), file=sys.stderr if a.pcm else sys.stdout) |
|
|
|
|
| if __name__ == "__main__": |
| main() |
|
|
