Synopsis
Description
Transcribes audio files to text using automatic speech recognition (ASR). Supports multiple audio formats and output options.Arguments
Options
The CLI uploads the audio file as multipart form data. The server accepts up to 25 MiB per file by default; raise it with
IZWI_OPENAI_AUDIO_UPLOAD_LIMIT_MB on the server.
The CLI waits at most 5 minutes for the transcript. For long recordings on a slower machine, use a transcription job in the web UI or /v1/speech-to-text/jobs instead.
Examples
Basic transcription
Save to file
JSON output
Prompt-guided Granite Speech transcription
Granite Speech 5.0 translation and keyword biasing
--translate-to and --keyword need a model with speech translation or
keyword biasing (Granite-Speech-5.0-2B, its GGUF release and
Nemotron-3.5-Transcribe-30B-A4B-FP8); the server rejects them for other
models.
Nemotron 3.5 Transcribe
Nemotron-3.5-Transcribe-30B-A4B-FP8 needs about 32 GiB of memory. It transcribes up to 60 minutes in one pass in 40 languages; without
--language, or with auto, the model detects the language. It times words
with its own CTC head, so --word-timestamps needs no aligner.
--speaker-labels and --keyword cannot be combined with --translate-to,
and a translation has no word timestamps. --prompt replaces the model’s
whole instruction, so it cannot be combined with --translate-to,
--keyword or --speaker-labels; the server refuses those requests.
A one-hour WAV is larger than the upload limit; send long recordings as MP3 or
M4A. See Transcription.
LFM3 Audio transcription
LFM3-Audio-2B-GGUF (and LFM3-Audio-2B-Q8_0-GGUF) transcribes 18
languages: en, ar, da, de, el, es, fi, fr, he, it, ja,
ko, nl, pl, pt, sv, tr and zh. Without --language, or with
auto, the model chooses; the server refuses any other code with 400. It
ignores --prompt, has no word timestamps of its own, and transcribes long
audio in chunks of at most 30 seconds.
Word timestamps
With timestamps
Specify language
Use larger model
Use Voxtral
Output Formats
Text
Plain text transcript:JSON
Verbose JSON
--word-timestamps requests verbose_json output with a words array when the
selected model or forced aligner can provide word timing metadata.
--speaker-labels requests verbose_json output with a speaker_turns array:
one entry per run of one speaker’s text, with the model’s speaker tag number
(null before the first tag) and start/end seconds when the model timed
the words. The text field stays untagged.
Supported Audio Formats
- WAV (
.wav) - MP3 (
.mp3) - M4A (
.m4a) - FLAC (
.flac) - OGG (
.ogg) - WebM (
.webm)