@juicesharp/rpiv-voice

Dictate long prompts to Pi Agent instead of typing them. rpiv-voice adds a /voice command: an overlay opens, you speak, you press Enter, and the transcript drops into Pi's editor. Speech-to-text runs on your own CPU through sherpa-onnx Whisper (base multilingual int8) — no cloud, no API key, no account.

Install

pi install npm:@juicesharp/rpiv-voice

Restart your Pi session.

Quick start

Type /voice. The first run downloads the Whisper model (~198 MB, ~157 MB on disk) into ~/.pi/models/whisper-base/ with a progress splash; later runs open in about a second.

Then talk. Committed text appears as you finish phrases, with a dim trailing partial for the sentence you are still speaking.

| Key | Action | | --- | --- | | Enter | Close the overlay and paste the transcript — committed text plus the dim partial | | Esc | Close the overlay and paste nothing | | Space | Pause / resume | | Tab | Open the settings screen |

If you have remapped Pi's confirm/cancel keys, the overlay follows your remap.

What you get

• Audio never leaves your machine — the only network call in the package is the one-time model download from GitHub Releases. Decoding runs locally on the CPU. No API key, no account, no telemetry.
• You see the words before you commit them — the still-open utterance is re-decoded about once a second and rendered dim after the committed text, and Enter pastes both. You never wait on a final decode.
• Long monologues stay responsive — pauses flush a segment, and a cap keeps a non-stop stretch from stalling the transcript.
• Built-in microphones work, not just headsets — capture falls back to the device's native sample rate when 16 kHz is refused.
• Whisper's silence hallucinations get filtered — a curated phrase set ("thanks for watching", "music", "applause", and non-English equivalents), a repetition-loop detector, and an input-side energy floor. Toggle it off when you are dictating single words.
• Localized UI in nine languages — de, en, es, fr, pt, pt-BR, ru, uk, zh. With @juicesharp/rpiv-i18n installed, /languages flips the overlay strings without a restart; without it the extension still loads, English-only.

Configuration

/voice needs no config file. The two toggles are editable on the settings screen (Tab from dictation, Ctrl-S to save, Esc to save silently and go back) and persist to ~/.config/rpiv-voice/voice.json; numThreads is JSON-only — the settings screen shows it read-only, so hand-edit the file and re-run /voice to apply a change:

| Key | Default | Effect | | --- | --- | --- | | hallucinationFilterEnabled | true | Drops Whisper's silence artifacts and repetition loops | | equalizerEnabled | false | Renders the live audio waveform under the transcript | | numThreads | 4 | Decode threads for the Whisper recognizer — integer 1–16; below 1 (or any invalid value) falls back to the default 4, above 16 clamps to 16. JSON-only, applied on the next /voice run |

The file is written with mode 0600.

The microphone is the OS default input and the model is fixed — neither is selectable.

Reference

• Configuration — every key, the XDG path rules, file permissions, the diagnostic log, and how the recognition language is chosen.
• Keys and screens — the full key table for both screens, the remappable bindings, and exactly what Enter pastes and Space pauses.
• Model install and platform support — the first-run pipeline, the files on disk, failure recovery, and the prebuilt-binary matrix.

Requirements

• Apple Silicon macOS, glibc Linux (x64 / arm64), or Windows x64. Intel Macs are not supported: the decibri capture library publishes no darwin-x64 binary. Alpine/musl is out for the same reason.
• A microphone Pi is allowed to use. On macOS, grant your terminal microphone access under System Settings → Privacy & Security → Microphone.
• tar with bzip2 support on PATH — model extraction shells out to tar -xjf.
• ~650 MB free during the first run — the archive and the unused fp32 weights are only deleted after extraction finishes. Settles at ~157 MB.
• Network access on the first run only. After that, /voice is fully offline.

Troubleshooting

• Microphone unavailable… error notification — the OS refused the input device. Grant your terminal microphone permission, confirm an input device is connected, then re-run /voice.
• /voice requires interactive mode — the command draws a TUI overlay and does nothing in a non-interactive session. Run /voice from an interactive Pi session.
• The transcript says "Thanks for watching" — Whisper hallucinated on near-silence. Speak closer to the microphone, and leave hallucinationFilterEnabled on.
• A phrase silently failed to transcribe — recognition errors are appended to ~/.config/rpiv-voice/errors.log rather than printed, because stderr would corrupt the live overlay. Check there for the underlying sherpa-onnx error.
• /voice is not found — restart your Pi session after installing.
• STT model files were removed or corrupted… — the install directory was damaged after a good download. It is wiped automatically; run /voice again to redownload.

Related

• @juicesharp/rpiv-i18n — optional. Install it with pi install npm:@juicesharp/rpiv-i18n to localize the overlay and enable /languages.
• @juicesharp/rpiv-pi — the umbrella package and /rpiv-setup. It does not install rpiv-voice; this package is opt-in and installed explicitly.

License

MIT — see LICENSE.