whisper

AI-powered speech recognition and transcription

TLDR

Transcribe audio file

$ whisper [audio.mp3]

Transcribe with specific model

$ whisper --model [medium] [audio.mp3]

Transcribe with language hint

$ whisper --language [en] [audio.mp3]

Output specific format

$ whisper --output_format [srt] [audio.mp3]

Translate to English

$ whisper --task translate [audio.mp3]

Output to specific directory

$ whisper --output_dir [/path/to/output] [audio.mp3]

Transcribe multiple files

$ whisper [audio1.mp3] [audio2.wav]

Use GPU with float16

$ whisper --device cuda --fp16 True [audio.mp3]

SYNOPSIS

whisper [--model size] [--language lang] [--task task] [--outputformat fmt] [options] files_

Whisper is OpenAI's automatic speech recognition (ASR) system. It transcribes audio in many languages and can translate to English.Model sizes trade accuracy for speed: tiny runs fastest, large is most accurate. The turbo model (default) offers a good balance, running ~8x faster than large with minor quality loss. The .en suffix (tiny.en, base.en) denotes English-only models, slightly better for English. The turbo model is not trained for translation tasks.Language detection is automatic but can be hinted. For non-English audio, specifying the language improves accuracy. Translation mode transcribes any language to English text.Output formats include plain text, subtitles (SRT, VTT), and JSON with timing data. Word-level timestamps enable karaoke-style highlighting.Processing uses GPU (CUDA) when available, significantly faster than CPU. The --fp16 flag enables half-precision math on compatible GPUs.Audio preprocessing handles various formats via FFmpeg. Long files are processed in segments with context maintained across segments.

PARAMETERS

--model SIZE

Model size: tiny, base, small, medium, large, turbo (default: turbo). English-only variants: tiny.en, base.en, small.en, medium.en.

--language LANG

Language code (en, de, fr, etc.) or auto.

--task TASK

Task: transcribe or translate.

--output_format FORMAT

Output format: txt, vtt, srt, tsv, json, all.

--output_dir DIR

Output directory.

--device DEVICE

Device: cpu, cuda.

--fp16 / --no-fp16

Use float16 (GPU) or float32.

--temperature TEMP

Sampling temperature.

--best_of NUM

Number of candidates.

--beam_size NUM

Beam search size.

--word_timestamps BOOL

Include word-level timestamps.

--condition_on_previous_text BOOL

Use previous output as context.

--verbose BOOL

Show progress and transcription.

--threads NUM

CPU threads.

--model_dir DIR

Directory to save and load models (default: ~/.cache/whisper).

--initial_prompt TEXT

Optional text to provide as prompt for the first window.

--clip_timestamps TIMESTAMPS

Comma-separated start/end timestamps to process specific audio segments.

CAVEATS

Large models require significant VRAM (10GB+ for large). CPU inference is slow. Accuracy varies by audio quality and accent. Hallucinations possible on silent or noisy segments. No speaker diarization. Model download required on first use.

HISTORY

Whisper was released by OpenAI in September 2022. Trained on 680,000 hours of multilingual audio, it achieved near-human transcription accuracy. The open-source release enabled local deployment, spawning community projects and integrations. The large-v3-turbo model was added in September 2024, offering significantly faster inference with minimal quality loss.