macOS Voice Toolkit — Pipe-based TTS + STT

Turn every built-in Mac voice into a Unix pipe filter. Zero disk I/O. Zero GPU. Pure CPU inference.

echo "Hello" | tts_pipe com.apple.voice.premium.en-US.Samantha 0.5 > out.pcm
echo "你好" | tts_pipe com.apple.voice.premium.zh-CN.Yun 0.5 | ffplay -f s16be -ar 22050 -ac 1 pipe:0 -

rec -t raw -r 16000 -c 1 -b 16 -e signed - | stt_pipe --lang zh-CN

tts_pipe — Text to Speech

Why not say?

say writes AIFF files (must seek, so pipes are impossible). tts_pipe calls AVSpeechSynthesizer.write(toBuffer:) to get PCM buffers directly in memory — a true Unix filter.

vs. Other Local TTS

Mac Premium voices (Yun, Han, Samantha, Kyoko, etc.) come from Apple's decades of speech synthesis R&D. Quality surpasses most open-source local TTS solutions. No GPU needed — pure CPU inference delivers natural, fluid speech.

Quick Start

make build-tts

# Play
./tts_pipe com.apple.voice.premium.zh-CN.Yun "Hello" 0.5 | ffplay -f s16be -ar 22050 -ac 1 pipe:0 -

# Install
sudo make install

Usage

# Single sentence
tts_pipe <voice-identifier> <text> [rate]

# Stream mode (stdin, one sentence per line)
tts_pipe --stdin <voice-identifier> [rate]

Recommended Voices

Chinese: Yun ★ Yue ★ Han ★ Lili ★ Lilian ★ Tingting English: Samantha ★ Alex ★ Ava ★ Japanese: Kyoko ★

★ = Premium (download in System Settings → Accessibility → Spoken Content)

List all installed:

swift -e 'import AVFoundation; AVSpeechSynthesisVoice.speechVoices().forEach { print($0.identifier, "|", $0.name, "|", $0.language) }'

Output Format

Sample rate 22050 Hz · 16-bit signed integer · mono · Big Endian

Usage from other languages

Python: subprocess.Popen(['tts_pipe', '--stdin', voice, rate], stdin=..., stdout=...) Node: spawn('tts_pipe', ['--stdin', voice, rate])

Full examples in examples/.

---

stt_pipe — Speech to Text

Why not use cloud APIs?

stt_pipe uses Apple's on-device SFSpeechRecognizer. All processing is local — no network, no API keys, no per-minute billing. On Apple Silicon, recognition runs on the Neural Engine.

Quick Start

make build-stt

# Pipe audio from a file
sox input.wav -t raw -r 16000 -c 1 -b 16 -e signed - | ./stt_pipe --lang zh-CN

# Record from microphone
./stt_pipe --mic --lang en-US

# With ffmpeg (any format)
ffmpeg -i recording.mp3 -f s16le -ar 16000 -ac 1 pipe:1 | ./stt_pipe --lang zh-CN

# Install
sudo make install

Usage

# Pipe mode (stdin → raw PCM)
stt_pipe --lang <locale> < audio.pcm

# Microphone mode
stt_pipe --mic --lang <locale>

Parameter	Description
--lang	Language locale (default: en-US). e.g. zh-CN, ja-JP, de-DE
--mic	Record from microphone instead of stdin

Supported Languages

60+ locales. Common ones:

Locale	Language
en-US	English (US)
zh-CN	Chinese (Simplified)
ja-JP	Japanese
ko-KR	Korean
de-DE	German
fr-FR	French
es-ES	Spanish

For the full list:

swift -e 'import Speech; SFSpeechRecognizer.supportedLocales().forEach { print($0.identifier) }' | sort

Audio Format (Pipe Mode)

Sample rate 16000 Hz · 16-bit signed integer · mono · Little Endian

Convert anything to this format with ffmpeg:

ffmpeg -i input.mp3 -f s16le -ar 16000 -ac 1 pipe:1 | stt_pipe --lang en-US

---

End-to-End: TTS → STT

Just for fun — synthesize with one voice and transcribe with another:

tts_pipe com.apple.voice.premium.zh-CN.Yun "你好世界" 0.5 2>/dev/null \
  | ffmpeg -f s16be -ar 22050 -ac 1 -i pipe:0 -f s16le -ar 16000 -ac 1 pipe:1 2>/dev/null \
  | stt_pipe --lang zh-CN 2>/dev/null
# → 你好世界

---

Install

git clone https://github.com/rrlab-tech/mac-tts.git
cd mac-tts
make build
sudo make install

This installs both tts_pipe and stt_pipe to /usr/local/bin/.

---

rrtts: Dart CLI Example

A higher-level Dart CLI that wraps tts_pipe with sentence splitting and WAV output:

cd rrtts && dart pub get
dart run bin/rrtts.dart -f README.md --voice=Yun

---

License

MIT.

Origin

Extracted from Juhoushu macOS voice modules.