Crate whisper_guard 

whisper-guard

The post-processing layer Whisper should have shipped with.

whisper.cpp and its bindings (whisper-rs, and many forks) hallucinate in predictable, well-documented ways: looping on silence, generating phantom [music] tags, drifting into a foreign script when the audio is too quiet, gluing voice commands like “stop recording” onto the end of every transcript.

whisper-guard catches the common patterns, with defaults tuned in production by Minutes, an OSS meeting-memory tool that processes meeting and voice-memo audio across multiple languages.

Quick start

If you already have Vec<String> segments from a transcription engine (whisper_state.get_segment(i).to_str(), a parakeet sidecar, a fork - anything):

use whisper_guard::clean_segments;

let raw = vec![
    "Thank you.".to_string(),
    "Thank you.".to_string(),
    "Thank you.".to_string(),
    "Thank you.".to_string(),
    "What's the budget for this quarter?".to_string(),
];

let (cleaned, stats) = clean_segments(&raw);
assert!(cleaned.iter().any(|s| s.contains("budget")));
println!("{}", stats.summary());
// → whisper-guard: 5 → 3 segments (2 removed)
//   (the loop collapses to first-occurrence + an annotation line; see CleanOptions)

That’s the whole API for the common case. No builders, no setup, no engine coupling. Six guards run in a fixed order; opt out individually via CleanOptions when you have a good reason.

Works with any whisper variant

whisper-guard’s segment/audio modules are pure Rust with no whisper-rs dependency. If you depend on a forked or pinned whisper-rs (common for Metal/CUDA tuning, GPU patches, or model compatibility), use:

whisper-guard = { version = "0.2", default-features = false }

…and the cleaning pipeline works regardless of which whisper-rs is in your tree. The optional whisper feature only adds params presets that wrap whisper_rs::FullParams.

What it catches

Pre-transcription audio prep (audio):

• Silence stripping with adaptive noise floor
• Auto-normalization for quiet microphones
• Windowed-sinc resampling (32-tap Hann, alias-free) for 44.1k → 16k

Post-transcription segment cleaning (segments):

• Consecutive repetition (3+ similar segments collapsed)
• Interleaved A/B/A/B hallucination patterns
• Bracketed noise marker collapse ([Śmiech], [music], [risas], any language)
• Foreign-script hallucination detection (e.g., CJK in a Latin transcript)
• Trailing noise trimming ([music], [BLANK_AUDIO], filler at the end)
• Voice command stripping (stop recording, end recording at the tail)

Whisper parameter presets (params, requires whisper feature):

• Batch transcription params matching whisper-cli defaults
• Low-latency streaming params

Re-exports

pub use audio::normalize_audio;

pub use audio::resample;

pub use audio::strip_silence;

pub use segments::clean_segments;

pub use segments::clean_segments_with_options;

pub use segments::clean_transcript;

pub use segments::strip_trailing_commands;

pub use segments::CleanOptions;

pub use segments::CleanStats;

Modules

audio

Pre-transcription audio preparation for whisper.

params

Whisper parameter presets for stable transcription.

segments

Post-transcription segment cleaning for whisper output.