Crate whisper_captions 

Turn whisper.cpp output into clean, caption-ready segments.

whisper.cpp can emit a JSON file (-oj / -ojf) or plain stdout lines like [00:00:01.000 --> 00:00:03.500] Hello there. This crate normalizes both into Segments with millisecond timestamps and, when token-level data is present, whole-word Word timings.

whisper.cpp’s special tokens ([_BEG_], [_TT_150], <|endoftext|>, …) are dropped, and to_srt / to_vtt turn segments into subtitle files.

No whisper dependency: it works on output files from any whisper.cpp version, and merge_bpe_tokens also accepts tokens from whisper-rs.

let stdout = "[00:00:01.000 --> 00:00:02.500]  Hello   world\n";
let segments = whisper_captions::parse_whisper_stdout(stdout);
assert_eq!(segments[0].text, "Hello world");
assert_eq!(segments[0].start_ms, 1000);
assert_eq!(segments[0].duration_ms, 1500);

let srt = whisper_captions::to_srt(&segments);
assert!(srt.starts_with("1\n00:00:01,000 --> 00:00:02,500\nHello world\n"));

Structs

RawToken

A raw whisper token before BPE merging. The text is kept untouched so the leading-space “starts a new word” signal survives until merging.

Segment

One caption-ready transcript segment.

Word

A whole word with its timing, reassembled from whisper’s BPE tokens.

Functions

is_special_token

Is this a whisper control token rather than speech?

merge_bpe_tokens

Reassemble whisper’s BPE tokens into whole words.

normalize_raw_token

Normalize one raw whisper word/token, keeping its text untouched.

parse_timestamp_ms

Parse a whisper timestamp into milliseconds.

parse_transcript

Prefer JSON segments; fall back to stdout when JSON is missing or empty.

parse_whisper_json

Parse a whisper.cpp JSON result (any of the known layouts).

parse_whisper_json_str

Parse a whisper.cpp JSON string. Invalid JSON yields an error.

parse_whisper_stdout

Parse whisper.cpp’s human-readable stdout ([start --> end] text lines).

to_srt

Render segments as SubRip (.srt).

to_vtt

Render segments as WebVTT (.vtt).