transcriber

Give it a URL or a file, get back a transcript with timestamps.

Built on whisper.cpp via whisper-rs. Handles the whole pipeline — downloading, audio decoding, resampling, transcription, output formatting — so you don't have to glue it together yourself.

How it works

graph LR
    A["URL or File"] --> B["yt-dlp
    1000+ sites"]
    A --> C

    B --> C["ffmpeg
    decode any format"]

    C --> D["compand + loudnorm
    compress & normalize"]

    D --> E["whisper.cpp
    CPU or GPU"]

    E --> F["Text"]
    E --> G["SRT"]
    E --> H["WebVTT"]
    E --> I["JSON"]

    style A fill:#2563eb,color:#fff,stroke:none
    style D fill:#7c3aed,color:#fff,stroke:none
    style E fill:#047857,color:#fff,stroke:none
    style F fill:#b45309,color:#fff,stroke:none
    style G fill:#b45309,color:#fff,stroke:none
    style H fill:#b45309,color:#fff,stroke:none
    style I fill:#b45309,color:#fff,stroke:none

Loading

Without transcriber you'd wire up yt-dlp, ffmpeg, sample rate conversion, and whisper yourself — different tools, different formats, lots of glue code. With transcriber it's one function call:

let transcript = transcriber::transcribe(url).await?;

Quick start

let transcript = transcriber::transcribe("https://youtube.com/watch?v=dQw4w9WgXcQ").await?;
println!("{}", transcript.text());

Or from a local file:

let transcript = transcriber::transcribe_file("meeting.mp3").await?;

// Get subtitles
println!("{}", transcript.to_srt());
println!("{}", transcript.to_vtt());

// Or structured data
let json = transcript.to_json_pretty()?;

Models are downloaded automatically from HuggingFace on first use and cached locally.

Install

[dependencies]
transcriber = "0.1"

For GPU acceleration:

[dependencies]
transcriber = { version = "0.1", features = ["cuda"] }
# or
transcriber = { version = "0.1", features = ["vulkan"] }

CLI

cargo install transcriber-cli

# Transcribe a YouTube video
transcriber-cli https://youtube.com/watch?v=... --format srt --output subtitles.srt

# Transcribe a local file
transcriber-cli recording.mp3 --model small --language de

# List available models
transcriber-cli --list-models

Options

use transcriber::{TranscribeOptions, Model};

let opts = TranscribeOptions::new()
    .model(Model::LargeV3Turbo)
    .language("de")?
    .word_timestamps(true)
    .translate(true)           // translate to English
    .gpu(true)
    .beam_size(5)?;

let transcript = transcriber::transcribe_with_options(url, &opts).await?;

for segment in &transcript.segments {
    println!("[{:.1}s - {:.1}s] {}", segment.start, segment.end, segment.text);
}

Audio conditioning

All audio is automatically conditioned before transcription via ffmpeg:

1. Dynamic range compression (compand) — boosts quiet speech (distant speakers, soft voices) while limiting loud peaks. Tuned for speech dynamics with 0.3s attack / 0.8s decay.
2. Loudness normalization (loudnorm) — EBU R128 normalization targeting -16 LUFS, the optimal input level for whisper.

This is critical for real-world recordings (meetings, lectures, interviews) where speakers are at different distances from the microphone.

Hallucination prevention

Whisper has a known failure mode where the decoder enters a repetition loop, generating the same phrase endlessly — especially on long recordings with quiet passages. transcriber prevents this at two levels:

• Decoder isolation: each 30-second window starts with a clean decoder slate (n_max_text_ctx=0), preventing hallucinated text from poisoning subsequent windows.
• Post-processing filter: a rolling-window detector catches repeated segments (exact repeats, alternating A/B patterns, short cycles) and removes them.

What it supports

Audio formats: anything ffmpeg can handle — mp3, wav, ogg, opus, flac, aac, m4a, webm, and more.

URL downloading: YouTube and 1000+ other sites via yt-dlp.

Output formats: plain text, SRT, WebVTT, JSON.

Models: tiny through large-v3-turbo, plus custom GGML files.

Languages: 100 languages with auto-detection.

Requirements

• ffmpeg — for audio decoding (apt install ffmpeg / brew install ffmpeg)
• yt-dlp — for URL downloads (pip install yt-dlp). Not needed if you only use local files — build without the download feature to drop this dependency.

Feature flags

Flag	Default	What it does
download	yes	URL downloading via yt-dlp
cuda	no	NVIDIA GPU acceleration
vulkan	no	Vulkan GPU acceleration

License

Licensed under either of Apache License, Version 2.0 or MIT license at your option.