Personal Voice To File

Personal Voice → file

Why this isn't a one-liner

Apple deliberately blocks AVSpeechSynthesizer's offline write(_:toBufferCallback:) API for Personal Voices — the callback never fires (confirmed: ordinary system voices export fine with the same code; only Personal Voice hangs). Signing, codesigning, and app-bundling do not change this. See NOTES.md for the full investigation.

The only working path: play the voice live through BlackHole (a virtual loopback audio device) instead of speakers, and capture that stream with ffmpeg.

One-time setup (per machine)

brew install blackhole-2ch switchaudio-osx
# BlackHole's installer needs an interactive sudo password — user must run this themselves:
sudo installer -pkg "$(brew --cache blackhole-2ch)" -target /
# Then either reboot, or try this first (sometimes enough, briefly cuts system audio):
sudo killall coreaudiod

Verify it worked: SwitchAudioSource -a -t output should list BlackHole 2ch.

Recording

scripts/record.sh <output.m4a> <voice-name-or-auto> <rate-multiplier> [text-file]

• voice-name-or-auto: substring match against installed Personal Voices (e.g. morgan), or auto.
• rate-multiplier: 1.0 = normal Apple speaking rate. talk-to-me's shared default is 1.25 — ask the user which they want, don't assume.
• text-file: optional; omit to read text from stdin.

The script: switches system audio output to BlackHole, records via ffmpeg -f avfoundation, runs personal-say (from the talk-to-me skill) with the given voice/rate, stops the capture, always restores the original output device (even on failure, via a trap), and transcodes the WAV to an AAC-in-MP4 file.

For long scripts (talk tracks, narration): strip markdown headers, stage directions in brackets, and horizontal rules before piping the text in — personal-say speaks raw text verbatim, punctuation and all.

Gotchas

• If SwitchAudioSource -c isn't "the device you expect" before you start, note it — the trap restores to whatever was captured at start, not a hardcoded default.
• Homebrew sometimes purges the downloaded .pkg from Caskroom/ after a failed install attempt. If the installer path 404s, re-run brew install blackhole-2ch (or brew fetch) and use the path it reports, or check ~/Library/Caches/Homebrew/downloads/.
• ffmpeg -f avfoundation -list_devices true -i "" re-numbers devices per machine — don't hardcode :0 for BlackHole; record.sh looks it up by name each run.
• talk-to-me's Stop hook will kill your recording mid-sentence. speak-response.sh runs a global pkill -x personal-say so a new response interrupts old narration — and it is not scoped to a session, so any Claude Code session ending a turn silently SIGTERMs the personal-say your capture is driving. The recording just stops (exit 143, Terminated: 15). Turning narration off does not help: other sessions still fire the hook. Fix: copy personal-say to a different filename (e.g. rec-say) and run the capture through that — pkill -x matches the exact process name, so a renamed copy is invisible to it. Turn this session's narration off too, or your own narration gets captured into the file.
• A long dialogue needs its own long settle times. BlackHole becoming the default output does not propagate to CoreAudio instantly; give it ~3s before starting ffmpeg and ~2s after, or personal-say renders to the old device and you capture pure silence. This failure is intermittent, which makes it read like a BlackHole problem when it is really a race.
• Don't inline the device lookup as a pipe into head. Under set -o pipefail, head closing early sends ffmpeg SIGPIPE and aborts the script with exit 251 and no output. Capture the device list into a variable first — record.sh already does this and says why.

Rate: the single biggest quality lever

TALKTOME_RATE is a multiplier on Apple's default, and Apple's default is already brisk. Anything at or above 1.15 makes Personal Voice slur words together and emit click artifacts — it sounds exactly like overlapping speech layered over a ticking background, which invites a long and wrong hunt through BlackHole, sample rates, and dropped buffers. Measured on this machine: 1.15 → ~277 wpm (unintelligible), 0.92 → ~197 wpm (clean).

For anything a human will actually sit and listen to, use 0.9–1.0. talk-to-me's 1.25 default is tuned for skimming live narration, not for a recorded file.

Verify a finished file rather than trusting it — decode to mono and check three things: the count of >0.7s silences matches the number of turns, words-per-minute is under ~210, and the loudest sample inside a turn gap is 0 (real ticking shows up there; speech plosives do not).