Skip to main content

inspect-robots-voice

Local spoken operator feedback for attended Inspect Robots evaluations. The plugin keeps the microphone open during a run, transcribes accepted utterances locally, and sends them through the same operator-message channel as typed console feedback. Parakeet TDT 0.6B v3 is the default; faster-whisper models remain selectable by model name or local path.

Voice input is feedback-only. Spoken words cannot end a trial or record a verdict. Silence and low-confidence transcription candidates produce no message.

Install

pip install inspect-robots-voice

PortAudio must be available for sounddevice on the machine connected to the microphone: sudo apt install libportaudio2 (Debian and Ubuntu), sudo dnf install portaudio (Fedora), or brew install portaudio (macOS).

Run

inspect-robots run --task my-task --policy agent --embodiment my-robot --voice

Voice mode requires an attended terminal. Repeat -V key=value to configure it:

Key Default Backend Meaning
model parakeet-tdt-0.6b-v3 both Parakeet alias or nemo- name, or a faster-whisper model size or local path
device system default both sounddevice input index or case-insensitive name substring
language none whisper explicit transcription language (Whisper uses en when unset); Parakeet auto-detects
compute auto whisper CTranslate2 compute type
asr_device cpu whisper cpu, cuda, or auto

The default int8 Parakeet weights download once from the Hugging Face hub on first use and use about 640 MB. Pass -V model=small to restore the previous Whisper default. Explicit language, compute, and non-CPU asr_device values require a Whisper model.

Parakeet TDT 0.6B v3 weights are provided by NVIDIA under the CC-BY-4.0 license.

For example:

inspect-robots run --task my-task --policy agent --embodiment my-robot \
    --voice -V model=medium -V device="USB Microphone" -V compute=int8

The plugin can also narrate streamed policy notes with run --speak. Speech defaults to interrupt, which cuts off superseded narration. Pass -S mode=blocking to wait boundedly for each previous note, or -S mode=queue to retain bounded drop-oldest queueing.

Filtering

Audio is segmented locally with an adaptive energy gate. Parakeet applies duration, blank-text, coarse mean-token-confidence, and known silence-hallucination checks. Whisper keeps its bundled VAD plus duration, confidence, no-speech, and known silence-hallucination checks. Rejected candidates disappear silently so the policy receives only accepted operator utterances.

Transcription runs in a worker thread. The rollout loop never waits for speech recognition, and a voice pipeline failure disables voice input without disabling typed console feedback.

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

inspect_robots_voice-0.5.1.tar.gz (32.1 kB view details)

Uploaded Source

Built Distribution

If you're not sure about the file name format, learn more about wheel file names.

inspect_robots_voice-0.5.1-py3-none-any.whl (19.2 kB view details)

Uploaded Python 3

File details

Details for the file inspect_robots_voice-0.5.1.tar.gz.

File metadata

  • Download URL: inspect_robots_voice-0.5.1.tar.gz
  • Upload date:
  • Size: 32.1 kB
  • Tags: Source
  • Uploaded using Trusted Publishing? Yes
  • Uploaded via: twine/7.0.0 CPython/3.13.14

File hashes

Hashes for inspect_robots_voice-0.5.1.tar.gz
Algorithm Hash digest
SHA256 1871f8dc8fcbef7300475fc47ee06d3f17d6bdead5eed4f036365c03d5f14b99
MD5 29c0c95aec115c290186a02d11026504
BLAKE2b-256 d6648cbe2b6f899f3d2d6bd422b1d93db7d300cdfee2e46409aeaaded728ccea

See more details on using hashes here.

Provenance

The following attestation bundles were made for inspect_robots_voice-0.5.1.tar.gz:

Publisher: release.yml on robocurve/inspect-robots

Attestations: Values shown here reflect the state when the release was signed and may no longer be current.

File details

Details for the file inspect_robots_voice-0.5.1-py3-none-any.whl.

File metadata

File hashes

Hashes for inspect_robots_voice-0.5.1-py3-none-any.whl
Algorithm Hash digest
SHA256 ef3dd25153d069d19b533d5973c9f1656076f99687c672c09407114b4c85bb50
MD5 53849cb41acdd984afdfe0b0622ab378
BLAKE2b-256 b0160ba29e19868b952506f3eddd98a6b1b56e05dca9c3b8710bd7ba933504d0

See more details on using hashes here.

Provenance

The following attestation bundles were made for inspect_robots_voice-0.5.1-py3-none-any.whl:

Publisher: release.yml on robocurve/inspect-robots

Attestations: Values shown here reflect the state when the release was signed and may no longer be current.

Release history Release notifications | RSS feed

This release

0.5.1 This release

2 files

0.5.0

2 files

0.4.0

2 files

0.3.0

2 files

0.2.0

2 files

0.1.0

2 files

Anthropic, PBC Visionary sponsor Bloomberg Visionary sponsor Hudson River Trading Visionary sponsor Meta Visionary sponsor NVIDIA Visionary sponsor Microsoft Sustainability sponsor Depot Continuous Integration AWS Cloud computing and Security Sponsor Datadog Monitoring Fastly CDN Google Download Analytics Sentry Error logging StatusPage Status page