Faster Whisper

Faster Whisper

Local speech-to-text using faster-whisper — a CTranslate2 reimplementation of OpenAI's Whisper that runs 4-6x faster with identical accuracy. With GPU acceleration, expect ~20x realtime transcription (a 10-minute audio file in ~30 seconds).

When to Use

Use this skill when you need to:

• Transcribe audio/video files — meetings, interviews, podcasts, lectures, YouTube videos
• Generate subtitles — SRT, VTT, ASS, LRC, or TTML broadcast-standard subtitles
• Identify speakers — diarization labels who said what (--diarize)
• Transcribe from URLs — YouTube links and direct audio URLs (auto-downloads via yt-dlp)
• Transcribe podcast feeds — --rss <feed-url> fetches and transcribes episodes
• Batch process files — glob patterns, directories, skip-existing support; ETA shown automatically
• Convert speech to text locally — no API costs, works offline (after model download)
• Translate to English — translate any language to English with --translate
• Do multilingual transcription — supports 99+ languages with auto-detection
• Transcribe a batch of files in different languages — --language-map assigns a different language per file
• Transcribe multilingual audio — --multilingual for mixed-language audio
• Transcribe audio with specific terms — use --initial-prompt for jargon-heavy content or any other terms to look out for
• Preprocess noisy audio (before transcription) — --normalize and --denoise before transcription
• Stream output — --stream shows segments as they're transcribed
• Clip time ranges — --clip-timestamps to transcribe specific sections
• Search the transcript — --search "term" finds all timestamps where a word/phrase appears
• Detect chapters — --detect-chapters finds section breaks from silence gaps
• Export speaker audio — --export-speakers DIR saves each speaker's turns as separate WAV files
• Spreadsheet output — --format csv produces a properly-quoted CSV with timestamps

Trigger phrases: "transcribe this audio", "convert speech to text", "what did they say", "make a transcript", "audio to text", "subtitle this video", "who's speaking", "translate this audio", "translate to English", "find where X is mentioned", "search transcript for", "when did they say", "at what timestamp", "add chapters", "detect chapters", "find breaks in the audio", "table of contents for this recording", "TTML subtitles", "DFXP subtitles", "broadcast format subtitles", "Netflix format", "ASS subtitles", "aegisub format", "advanced substation alpha", "mpv subtitles", "LRC subtitles", "timed lyrics", "karaoke subtitles", "music player lyrics", "HTML transcript", "confidence-colored transcript", "color-coded transcript", "separate audio per speaker", "export speaker audio", "split by speaker", "transcript as CSV", "spreadsheet output", "transcribe podcast", "podcast RSS feed", "different languages in batch", "per-file language", "transcribe in multiple formats", "srt and txt at the same time", "output both srt and text", "remove filler words", "clean up ums and uhs", "strip hesitation sounds", "remove you know and I mean", "transcribe left channel", "transcribe right channel", "stereo channel", "left track only", "wrap subtitle lines", "character limit per line", "max chars per subtitle", "detect paragraphs", "paragraph breaks", "group into paragraphs", "add paragraph spacing"

⚠️ Agent guidance — keep invocations minimal:

CORE RULE: default command (./scripts/transcribe audio.mp3) is the fastest path — add flags only when the user explicitly asks for that capability.

Transcription:

• Only add --diarize if the user asks "who said what" / "identify speakers" / "label speakers"
• Only add --format srt/vtt/ass/lrc/ttml if the user asks for subtitles/captions in that format
• Only add --format csv if the user asks for CSV or spreadsheet output
• Only add --word-timestamps if the user needs word-level timing
• Only add --initial-prompt if there's domain-specific jargon to prime
• Only add --translate if the user wants non-English audio translated to English
• Only add --normalize/--denoise if the user mentions bad audio quality or noise
• Only add --stream if the user wants live/progressive output for long files
• Only add --clip-timestamps if the user wants a specific time range
• Only add --temperature 0.0 if the model is hallucinating on music/silence
• Only add --vad-threshold if VAD is aggressively cutting speech or including noise
• Only add --min-speakers/--max-speakers when you know the speaker count
• Only add --hf-token if the token is not cached at ~/.cache/huggingface/token
• Only add --max-words-per-line for subtitle readability on long segments
• Only add --filter-hallucinations if the transcript contains obvious artifacts (music markers, duplicates)
• Only add --merge-sentences if the user asks for sentence-level subtitle cues
• Only add --clean-filler if the user asks to remove filler words (um, uh, you know, I mean, hesitation sounds)
• Only add --channel left|right if the user mentions stereo tracks, dual-channel recordings, or asks for a specific channel
• Only add --max-chars-per-line N when the user specifies a character limit per subtitle line (e.g., "Netflix format", "42 chars per line"); takes priority over --max-words-per-line
• Only add --detect-paragraphs if the user asks for paragraph breaks or structured text output; --paragraph-gap (default 3.0s) only if they want a custom gap
• Only add --speaker-names "Alice,Bob" when the user provides real names to replace SPEAKER_1/2 — always requires --diarize
• Only add --hotwords WORDS when the user names specific rare terms not well served by --initial-prompt; prefer --initial-prompt for general domain jargon
• Only add --prefix TEXT when the user knows the exact words the audio starts with
• Only add --detect-language-only when the user only wants to identify the language, not transcribe
• Only add --stats-file PATH if the user asks for performance stats, RTF, or benchmark info
• Only add --parallel N for large CPU batch jobs; GPU handles one file efficiently on its own — don't add for single files or small batches
• Only add --retries N for unreliable inputs (URLs, network files) where transient failures are expected
• Only add --burn-in OUTPUT only when user explicitly asks to embed/burn subtitles into the video; requires ffmpeg and a video file input
• Only add --keep-temp when the user may re-process the same URL to avoid re-downloading
• Only add --output-template when user specifies a custom naming pattern in batch mode
• Multi-format output (--format srt,text): only when user explicitly wants multiple formats in one pass; always pair with -o <dir>
• Any word-level feature auto-runs wav2vec2 alignment (~5-10s overhead)
• --diarize adds ~20-30s on top of that

Search:

• Only add --search "term" when the user asks to find/locate/search for a specific word or phrase in audio
• --search replaces the normal transcript output — it prints only matching segments with timestamps
• Add --search-fuzzy only when the user mentions approximate/partial matching or typos
• To save search results to a file, use -o results.txt

Chapter detection:

• Only add --detect-chapters when the user asks for chapters, sections, a table of contents, or "where does the topic change"
• Default --chapter-gap 8 (8-second silence = new chapter) works for most podcasts/lectures; tune down for dense content
• --chapter-format youtube (default) outputs YouTube-ready timestamps; use json for programmatic use
• Always use --chapters-file PATH when combining chapters with a transcript output — avoids mixing chapter markers into the transcript text
• If the user only wants chapters (not the transcript), pipe stdout to a file with -o /dev/null and use --chapters-file
• Batch mode limitation: --chapters-file takes a single path — in batch mode, each file's chapters overwrite the previous. For batch chapter detection, omit --chapters-file (chapters print to stdout under === CHAPTERS (N) ===) or use a separate run per file

Speaker audio export:

• Only add --export-speakers DIR when the user explicitly asks to save each speaker's audio separately
• Always pair with --diarize — it silently skips if no speaker labels are present
• Requires ffmpeg; outputs SPEAKER_1.wav, SPEAKER_2.wav, etc. (or real names if --speaker-names is set)

Language map:

• Only add --language-map in batch mode when the user has confirmed different languages across files
• Inline format: "interview*.mp3=en,lecture*.mp3=fr" — fnmatch globs on filename
• JSON file format: @/path/to/map.json where the file is {"pattern": "lang_code"}

RSS / Podcast:

• Only add --rss URL when the user provides a podcast RSS feed URL
• Default fetches 5 newest episodes; --rss-latest 0 for all; --skip-existing to resume safely
• Always use -o <dir> with --rss — without it, all episode transcripts print to stdout concatenated, which is hard to use; each episode gets its own file when -o <dir> is set

Output format for agent relay:

• Search results (--search) → print directly to user; output is human-readable
• Chapter output → if no --chapters-file, chapters appear in stdout under === CHAPTERS (N) === header after the transcript; with --format json, chapters are also embedded in the JSON under "chapters" key
• Subtitle formats (SRT, VTT, ASS, LRC, TTML) → always write to -o file; tell the user the output path, never paste raw subtitle content
• Data formats (CSV, HTML, TTML, JSON) → always write to -o file; tell the user the output path, don't paste raw XML/CSV/HTML
• ASS format → for Aegisub, VLC, mpv; write to file and tell user they can open it in Aegisub or play it in VLC/mpv
• LRC format → timed lyrics for music players (Foobar2000, AIMP, VLC); write to file
• Multi-format (--format srt,text) → requires -o <dir>; each format goes to a separate file; tell user all paths written
• JSON format → useful for programmatic post-processing; not ideal to paste in full to user
• Text/transcript → safe to show directly to user for short files; summarise for long ones
• Stats output (--stats-file) → summarise key fields (duration, processing time, RTF) for the user rather than pasting raw JSON
• Language detection (--detect-language-only) → print the result directly; it's a single line
• ETA is printed automatically to stderr for batch jobs; no action needed

When NOT to use:

• Cloud-only environments without local compute
• Files <10 seconds where API call latency doesn't matter

faster-whisper vs whisperx: This skill covers everything whisperx does — diarization (--diarize), word-level timestamps (--word-timestamps), SRT/VTT subtitles — so whisperx is not needed. Use whisperx only if you specifically need its pyannote pipeline or batch-GPU features not covered here.

Quick Reference

Model Selection

Choose the right model for your needs:

digraph model_selection {
    rankdir=LR;
    node [shape=box, style=rounded];

    start [label="Start", shape=doublecircle];
    need_accuracy [label="Need maximum\naccuracy?", shape=diamond];
    multilingual [label="Multilingual\ncontent?", shape=diamond];
    resource_constrained [label="Resource\nconstraints?", shape=diamond];

    large_v3 [label="large-v3\nor\nlarge-v3-turbo", style="rounded,filled", fillcolor=lightblue];
    large_turbo [label="large-v3-turbo", style="rounded,filled", fillcolor=lightblue];
    distil_large [label="distil-large-v3.5\n(default)", style="rounded,filled", fillcolor=lightgreen];
    distil_medium [label="distil-medium.en", style="rounded,filled", fillcolor=lightyellow];
    distil_small [label="distil-small.en", style="rounded,filled", fillcolor=lightyellow];

    start -> need_accuracy;
    need_accuracy -> large_v3 [label="yes"];
    need_accuracy -> multilingual [label="no"];
    multilingual -> large_turbo [label="yes"];
    multilingual -> resource_constrained [label="no (English)"];
    resource_constrained -> distil_small [label="mobile/edge"];
    resource_constrained -> distil_medium [label="some limits"];
    resource_constrained -> distil_large [label="no"];
}

Model Table

Standard Models (Full Whisper)

Distilled Models (~6x Faster, ~1% WER difference)

.en models are English-only and slightly faster/better for English content.

Note for distil models: HuggingFace recommends disabling condition_on_previous_text for all distil models to prevent repetition loops. The script auto-applies --no-condition-on-previous-text whenever a distil-* model is detected. Pass --condition-on-previous-text to override if needed.

Custom & Fine-tuned Models

WhisperModel accepts local CTranslate2 model directories and HuggingFace repo names — no code changes needed.

Load a local CTranslate2 model

./scripts/transcribe audio.mp3 --model /path/to/my-model-ct2

Convert a HuggingFace model to CTranslate2

pip install ctranslate2
ct2-transformers-converter \
  --model openai/whisper-large-v3 \
  --output_dir whisper-large-v3-ct2 \
  --copy_files tokenizer.json preprocessor_config.json \
  --quantization float16
./scripts/transcribe audio.mp3 --model ./whisper-large-v3-ct2

Load a model by HuggingFace repo name (auto-downloads)

./scripts/transcribe audio.mp3 --model username/whisper-large-v3-ct2

Custom model cache directory

By default, models are cached in ~/.cache/huggingface/. Use --model-dir to override:

./scripts/transcribe audio.mp3 --model-dir ~/my-models

Setup

Linux / macOS / WSL2

# Base install (creates venv, installs deps, auto-detects GPU)
./setup.sh

# With speaker diarization support
./setup.sh --diarize

Requirements:

• Python 3.10+
• ffmpeg is not required for basic transcription — PyAV (bundled with faster-whisper) handles audio decoding. ffmpeg is only needed for --burn-in, --normalize, and --denoise.
• Optional: yt-dlp (for URL/YouTube input)
• Optional: pyannote.audio (for --diarize, installed via setup.sh --diarize)

Platform Support

*faster-whisper uses CTranslate2 which is CPU-only on macOS, but Apple Silicon is fast enough for practical use.

GPU Support (IMPORTANT!)

The setup script auto-detects your GPU and installs PyTorch with CUDA. Always use GPU if available — CPU transcription is extremely slow.

RTX 3070 tip: Use --compute-type int8_float16 for hybrid quantization — saves ~1GB VRAM with minimal quality loss. Ideal for running diarization alongside transcription.

If setup didn't detect your GPU, manually install PyTorch with CUDA:

# For CUDA 12.x
uv pip install --python .venv/bin/python torch --index-url https://download.pytorch.org/whl/cu121

# For CUDA 11.x
uv pip install --python .venv/bin/python torch --index-url https://download.pytorch.org/whl/cu118

• WSL2 users: Ensure you have the NVIDIA CUDA drivers for WSL installed on Windows

Usage

# Basic transcription
./scripts/transcribe audio.mp3

# SRT subtitles
./scripts/transcribe audio.mp3 --format srt -o subtitles.srt

# WebVTT subtitles
./scripts/transcribe audio.mp3 --format vtt -o subtitles.vtt

# Transcribe from YouTube URL
./scripts/transcribe https://youtube.com/watch?v=dQw4w9WgXcQ --language en

# Speaker diarization
./scripts/transcribe meeting.wav --diarize

# Diarized VTT subtitles
./scripts/transcribe meeting.wav --diarize --format vtt -o meeting.vtt

# Prime with domain terminology
./scripts/transcribe lecture.mp3 --initial-prompt "Kubernetes, gRPC, PostgreSQL, NGINX"

# Batch process a directory
./scripts/transcribe ./recordings/ -o ./transcripts/

# Batch with glob, skip already-done files
./scripts/transcribe *.mp3 --skip-existing -o ./transcripts/

# Filter low-confidence segments
./scripts/transcribe noisy-audio.mp3 --min-confidence 0.6

# JSON output with full metadata
./scripts/transcribe audio.mp3 --format json -o result.json

# Specify language (faster than auto-detect)
./scripts/transcribe audio.mp3 --language en

Options

Input:
  AUDIO                 Audio file(s), directory, glob pattern, or URL
                        Accepts: mp3, wav, m4a, flac, ogg, webm, mp4, mkv, avi, wma, aac
                        URLs auto-download via yt-dlp (YouTube, direct links, etc.)

Model & Language:
  -m, --model NAME      Whisper model (default: distil-large-v3.5; "turbo" = large-v3-turbo)
  --revision REV        Model revision (git branch/tag/commit) to pin a specific version
  -l, --language CODE   Language code, e.g. en, es, fr (auto-detects if omitted)
  --initial-prompt TEXT  Prompt to condition the model (terminology, formatting style)
  --prefix TEXT         Prefix to condition the first segment (e.g. known starting words)
  --hotwords WORDS      Space-separated hotwords to boost recognition
  --translate           Translate any language to English (instead of transcribing)
  --multilingual        Enable multilingual/code-switching mode (helps smaller models)
  --hf-token TOKEN      HuggingFace token for private/gated models and diarization
  --model-dir PATH      Custom model cache directory (default: ~/.cache/huggingface/)

Output Format:
  -f, --format FMT      text | json | srt | vtt | tsv | lrc | html | ass | ttml (default: text)
                        Accepts comma-separated list: --format srt,text writes both in one pass
                        Multi-format requires -o <dir> when saving to files
  --word-timestamps     Include word-level timestamps (wav2vec2 aligned automatically)
  --stream              Output segments as they are transcribed (disables diarize/alignment)
  --max-words-per-line N  For SRT/VTT, split segments into sub-cues of at most N words
  --max-chars-per-line N  For SRT/VTT/ASS/TTML, split lines so each fits within N characters
                        Takes priority over --max-words-per-line when both are set
  --clean-filler        Remove hesitation fillers (um, uh, er, ah, hmm, hm) and discourse markers
                        (you know, I mean, you see) from transcript text. Off by default.
  --detect-paragraphs   Insert paragraph breaks (blank lines) in text output at natural boundaries.
                        A new paragraph starts when: silence gap ≥ --paragraph-gap, OR the previous
                        segment ends a sentence AND the gap ≥ 1.5s.
  --paragraph-gap SEC   Minimum silence gap in seconds to start a new paragraph (default: 3.0).
                        Used with --detect-paragraphs.
  --channel {left,right,mix}
                        Stereo channel to transcribe: left (c0), right (c1), or mix (default: mix).
                        Extracts the channel via ffmpeg before transcription. Requires ffmpeg.
  --merge-sentences     Merge consecutive segments into sentence-level chunks
                        (improves SRT/VTT readability; groups by terminal punctuation or >2s gap)
  -o, --output PATH     Output file or directory (directory for batch mode)
  --output-template TEMPLATE
                        Batch output filename template. Variables: {stem}, {lang}, {ext}, {model}
                        Example: "{stem}_{lang}.{ext}" → "interview_en.srt"

Inference Tuning:
  --beam-size N         Beam search size; higher = more accurate but slower (default: 5)
  --temperature T       Sampling temperature or comma-separated fallback list, e.g.
                        '0.0' or '0.0,0.2,0.4' (default: faster-whisper's schedule)
  --no-speech-threshold PROB
                        Probability threshold to mark segments as silence (default: 0.6)
  --batch-size N        Batched inference batch size (default: 8; reduce if OOM)
  --no-vad              Disable voice activity detection (on by default)
  --vad-threshold T     VAD speech probability threshold (default: 0.5)
  --vad-neg-threshold T VAD negative threshold for ending speech (default: auto)
  --vad-onset T         Alias for --vad-threshold (legacy)
  --vad-offset T        Alias for --vad-neg-threshold (legacy)
  --min-speech-duration MS  Minimum speech segment duration in ms (default: 0)
  --max-speech-duration SEC Maximum speech segment duration in seconds (default: unlimited)
  --min-silence-duration MS Minimum silence before splitting a segment in ms (default: 2000)
  --speech-pad MS       Padding around speech segments in ms (default: 400)
  --no-batch            Disable batched inference (use standard WhisperModel)
  --hallucination-silence-threshold SEC
                        Skip silent sections where model hallucinates (e.g. 1.0)
  --no-condition-on-previous-text
                        Don't condition on previous text (reduces repetition/hallucination loops;
                        auto-enabled for distil models per HuggingFace recommendation)
  --condition-on-previous-text
                        Force-enable conditioning on previous text (overrides auto-disable for distil models)
  --compression-ratio-threshold RATIO
                        Filter segments above this compression ratio (default: 2.4)
  --log-prob-threshold PROB
                        Filter segments below this avg log probability (default: -1.0)
  --max-new-tokens N    Maximum tokens per segment (prevents runaway generation)
  --clip-timestamps RANGE
                        Transcribe specific time ranges: '30,60' or '0,30;60,90' (seconds)
  --progress            Show transcription progress bar
  --best-of N           Candidates when sampling with non-zero temperature (default: 5)
  --patience F          Beam search patience factor (default: 1.0)
  --repetition-penalty F  Penalty for repeated tokens (default: 1.0)
  --no-repeat-ngram-size N  Prevent n-gram repetitions of this size (default: 0 = off)

Advanced Inference:
  --no-timestamps       Output text without timing info (faster; incompatible with
                        --word-timestamps, --format srt/vtt/tsv, --diarize)
  --chunk-length N      Audio chunk length in seconds for batched inference (default: auto)
  --language-detection-threshold T
                        Confidence threshold for language auto-detection (default: 0.5)
  --language-detection-segments N
                        Audio segments to sample for language detection (default: 1)
  --length-penalty F    Beam search length penalty; >1 favors longer, <1 favors shorter (default: 1.0)
  --prompt-reset-on-temperature T
                        Reset initial prompt when temperature fallback hits threshold (default: 0.5)
  --no-suppress-blank   Disable blank token suppression (may help soft/quiet speech)
 

…(truncated)