Pause / Delay between paragraphs

[AeGhBrA]

Hello, Is there any way, either through a command or through text editing, to add different lengths of pause between the paragraphs (and possibly after "," or ".")? I've tried adding multiple ellipsis and other punctuation and symbols, but it either ignores multiple instances or reads them out loud in some cases.

_{Personally, I'm using en-US-BrianNeural (because it sounds the most natural to me), but it doesn't really care much about punctuation apparently, and will fire out all the words quicker than I can absorb them, since I mainly use this to turn various webnovels into audio while I do other things.}

[wmorin]

Not with markup, unfortunately. edge-tts escapes the text you give it (escape(...) in communicate.py) and wraps it in a single <prosody> tag, so a <break time="1s"/> would just be read out. The README explains why: Microsoft only accepts the SSML Edge itself generates, so custom SSML support was removed.

Punctuation helps a little but isn't controllable. With BrianNeural I measured about 0.45 s after a plain . (blank lines change nothing), 0.65 s with . ... and 1.3 s with . ... ... ... , but adding even more dots made the pause shorter again.

What works reliably is what rany2 suggested in #83: synthesize one paragraph at a time and put real silence in between. A small script for that (needs pip install edge-tts and ffmpeg):

import asyncio, pathlib, re, subprocess, sys, tempfile
import edge_tts

VOICE = "en-US-BrianNeural"
RATE = "-10%"      # a bit slower overall
PAUSE = 1.2        # seconds of silence between paragraphs

async def main(src, out):
    text = pathlib.Path(src).read_text(encoding="utf-8")
    # blank-line separated paragraphs if there are any, otherwise one paragraph per line
    blocks = re.split(r"\n\s*\n", text) if re.search(r"\n\s*\n", text) else text.splitlines()
    paras = [p.strip() for p in blocks if p.strip()]
    tmp = pathlib.Path(tempfile.mkdtemp())
    silence = tmp / "silence.mp3"
    subprocess.run(["ffmpeg", "-loglevel", "error", "-f", "lavfi", "-i", "anullsrc=r=24000:cl=mono",
                    "-t", str(PAUSE), "-c:a", "libmp3lame", "-b:a", "48k", str(silence)], check=True)
    parts = []
    for i, p in enumerate(paras):
        f = tmp / f"p{i:04d}.mp3"
        await edge_tts.Communicate(p, VOICE, rate=RATE).save(str(f))
        parts += [f, silence]
    listfile = tmp / "list.txt"
    listfile.write_text("".join(f"file '{x}'\n" for x in parts[:-1]))
    subprocess.run(["ffmpeg", "-loglevel", "error", "-y", "-f", "concat", "-safe", "0",
                    "-i", str(listfile), "-c:a", "libmp3lame", "-b:a", "48k", out], check=True)

asyncio.run(main(sys.argv[1], sys.argv[2]))

python paragraphs_tts.py chapter.txt chapter.mp3. It works whether your paragraphs are separated by blank lines or are one per line. On a short sample I got gaps of about 1.8 s (the 1.2 s plus the voice's own trailing silence); change PAUSE to taste. If it also reads too fast, RATE maps to the CLI's --rate=-10% (the = is needed there because of the minus sign).

[AeGhBrA]

Wow, it works really nice. The . ... ... ... trick also kinda works, but it's juuuust a little short (by about 300ms) and it's funky when, occasionally, lines don't end in full stops, so just ... ... ... \r\n will actually be shorter for some reason; the script, on the other hand, works great!

I was afraid it might not work with my AHK GUI script (that takes all dropped files and sends them to edge-tts to process simultaneously in the background) since I didn't know what it does with tmp files, but it seems to not mind at all; only my edge-tts.exe tracker (which just pings me when they're all done) is now a bit wonky, since it doesn't directly open edge-tts.exe, but it's all good.

Thank you, very much!