Text-to-Speech

Use pocket-tts for local text-to-speech. It is installed at ~/.local/bin/pocket-tts and is usually on PATH. Pass -q to keep logs quiet.

Start with the CLI help when details matter:

pocket-tts --help
pocket-tts generate --help
pocket-tts export-voice --help

Speak Aloud

Generate into a temporary directory, play it, then remove it:

d=$(mktemp -d /tmp/tts_XXXXXX) && f="$d/out.wav" && pocket-tts generate --text "Your message here" -q --output-path "$f" && afplay "$f"; rm -rf "$d"

Use a temp directory because pocket-tts writes WAV output and macOS mktemp does not safely randomize templates like /tmp/tts_XXXX.wav.

Save Audio

When the user asks for a file, write the WAV to the requested path:

pocket-tts generate --text "Chapter one. It was a bright cold day in April." -q --output-path ~/Desktop/narration.wav

Output is WAV: mono, 24 kHz, 16-bit PCM.

Languages

Use --language for non-English text. The English model will pronounce German and other languages badly.

pocket-tts generate --text "Guten Morgen, wie geht es dir?" --language german_24l -q --output-path "$f"

Useful language values: english, french24l, german24l, spanish24l, portuguese24l, italian24l. The 24l models are larger preview models; first use may take a minute or two while weights download or load.

Voices

Default English voice is alba. Non-English defaults are selected by language: estelle for French, juergen for German, lola for Spanish, rafael for Portuguese, and giovanni for Italian.

Use --voice only when the user asks for a specific voice:

pocket-tts generate --text "Hello there." --voice paul -q --output-path "$f"

Examples of built-in voices: alba, anna, vera, mary, jane, jean, paul, george, michael. Run pocket-tts generate --help or try a short sample if you need to compare options.

Voice Cloning

For a one-off custom voice from a clean 5-30 second sample:

pocket-tts generate --text "Hello" --voice /path/to/sample.wav -q --output-path "$f"

For reuse:

pocket-tts export-voice /path/to/sample.wav ~/voices/myvoice.safetensors -q
pocket-tts generate --text "Hello" --voice ~/voices/myvoice.safetensors -q --output-path "$f"

Voice cloning can require one-time Hugging Face access to kyutai/pocket-tts. If it fails with a gated-model message, tell the user to accept the model terms and run uvx hf auth login.

Practical Options

• --temperature: adjust variation if the result sounds too flat or unstable.
• --max-tokens: change chunk size for longer text if needed.
• --quantize: reduce memory usage.
• pocket-tts serve: keep the model warm for many generations in a row.