Overview
Izwi’s transcription feature converts spoken audio into written text. Capabilities include:- High accuracy — State-of-the-art speech recognition
- Multiple formats — Support for WAV, MP3, M4A, FLAC, and more
- Language detection — Automatic language identification
- Timestamps — Optional word-level timing through a forced aligner, or from the model itself (Nemotron 3.5 Transcribe)
- Summaries — Optional AI summaries for persisted speech-text jobs
- Local processing — Complete privacy, no cloud
Getting Started
Download an ASR Model
Transcribe Audio
Using the CLI
Basic Usage
Options
The CLI uploads the file as multipart form data. The server accepts up to 25 MiB per file by default; set
IZWI_OPENAI_AUDIO_UPLOAD_LIMIT_MB on the server to raise it.
Examples
Basic transcription:Using the Web UI
- Navigate to Transcription in the sidebar
- Choose a mode
- Upload an audio file or record directly
- Select the required model
- Click Transcribe or submit the selected workflow
- View, copy, summarize, or download the result
Features
- Drag and drop — Upload files easily
- Record — Transcribe directly from microphone
- Mode switch — Choose Transcription, Speaker Attributed ASR, or Diarization
- Summaries — Generate or regenerate AI summaries for saved records
- Copy — One-click copy to clipboard
- Download — Save as text or JSON
Using the API
Endpoint
Request (multipart/form-data)
Example (curl)
Response (JSON)
Response (verbose_json)
type values such as
transcript.text.delta, transcript.text.done, and error.
See the API Reference for JSON input,
streaming events, upload limits, and exact response shapes.
Persisted Speech-Text Jobs
The web UI uses/v1/speech-to-text/jobs for saved transcription records,
Speaker Attributed ASR records, and diarization jobs:
job_kind=speaker_attributed_asr or job_kind=saa for speaker-turn
transcripts (Granite Speech 4.1 Plus or Nemotron 3.5 Transcribe), and
job_kind=diarization for speaker timelines.
Nemotron 3.5 Transcribe
Nemotron-3.5-Transcribe-30B-A4B-FP8 is NVIDIA’s 30B speech language model.
It transcribes a recording of up to 60 minutes in one pass, translates,
takes context terms, labels speakers and times words itself. It translates
short speech best, so a translation of a recording longer than 30 seconds
runs in segments cut at pauses and comes back joined in order.
It needs about 31 GiB of memory on Metal and CUDA and about 33 GiB on CPU;
a machine that cannot hold it is refused when the model loads.
Languages and wording
It transcribes 40 languages: Arabic, Bengali, Bulgarian, Chinese, Croatian, Czech, Danish, Dutch, English, Estonian, Finnish, French, German, Greek, Hebrew, Hindi, Hungarian, Indonesian, Italian, Japanese, Korean, Latvian, Lithuanian, Maltese, Marathi, Polish, Portuguese, Romanian, Russian, Slovak, Slovenian, Spanish, Swedish, Tagalog, Tamil, Telugu, Thai, Turkish, Ukrainian and Vietnamese. Name the language with a code (de, pt-BR) or an English
name, or leave it out (or send auto) to let the model detect it. Any other
language, Norwegian included, returns 400.
By default the transcript is the model card’s “readable text”: punctuation and
capitalization, with fillers and accidental repetitions left out. A prompt
(--prompt) replaces that instruction entirely, after tags such as
<spk:0> are removed (at most 4,000 characters). Because it replaces the
whole instruction, a prompt cannot be combined with translation, keywords
or speaker labels; such a request returns 400.
Translation
target_language on the API, --translate-to in the CLI, or Translate
to in the UI translates the speech into another language. Translation runs
to or from English only. Every language above translates except Bengali,
Indonesian, Marathi, Tagalog, Tamil, Telugu, Turkish and Vietnamese. A
translation has no word timestamps, and it cannot be combined with keywords or
speaker labels (the request returns 400).
Context terms
keywords on the API, --keyword in the CLI, or Keywords in the UI lists
names and vocabulary the speech contains. The model receives them as the
card’s “Context terms” instruction. Up to 64 terms of at most 80 characters
each; tags in a term are removed.
Speaker labels
The model tags speakers itself. On/v1/audio/transcriptions, set
speaker_labels=true with response_format=verbose_json (CLI:
--speaker-labels). The text field stays plain; the speaker turns come back
in speaker_turns:
speaker is the model’s tag number, or null for text before the first tag.
start and end are in seconds and appear only when the model timed the
words. Speaker labels work in Arabic, Chinese, English, French, German, Hindi,
Italian, Japanese, Korean, Portuguese, Russian, Spanish and Thai; naming
another language returns 400.
In the web UI, choose Speaker Attributed ASR and select this model. Its
turns are labelled “Speaker 1”, “Speaker 2” and so on, and timed when its word
timestamps align. The speaker expectation does not steer the model: a count
outside it only adds a warning.
For “who spoke when” segments, the model can also diarize on its own: name it
as the model of a diarization job (see
Single-Model Diarization).
It cannot be the ASR model of a Sortformer or Nemotron-3-Diarization pipeline.
Word timestamps
The model times words with its own CTC head, so no forced aligner is needed:timestamp_granularities[]=word on the API, --word-timestamps in the CLI,
or Include timestamps in the UI (no aligner model is asked for). If the
model cannot align a transcription job’s words, the server times them with the
forced aligner instead. Thai is split into words at spaces only, so Thai is
timed in larger pieces than NVIDIA’s own tools, which segment Thai into words.
Long recordings
The whole recording runs in one pass, up to 60 minutes; Izwi does not split it. Uploads keep their usual limits: 64 MiB for transcription jobs (the web UI), and 25 MiB by default on/v1/audio/transcriptions and in the CLI
(IZWI_OPENAI_AUDIO_UPLOAD_LIMIT_MB raises it). A one-hour WAV at 16 kHz,
mono, 16-bit is about 115 MB and is refused with 413. Send long recordings
compressed:
- An hour of mono MP3 at 48 kbit/s is about 22 MB and fits both limits. At 64 kbit/s (about 29 MB) it fits the job limit but not the default 25 MiB.
- FLAC is lossless but compresses speech much less; a full hour may still be close to 64 MiB.
/v1/speech-to-text/jobs.
izwi_asr_diagnostics.finish_reason says why decoding stopped: stop (the
model finished), length (the token budget ran out) or repetition (the
model repeated itself and was stopped).
Not supported
- Live transcription and voice mode: the model is offline only.
- The ASR stage of a diarization pipeline. It diarizes on its own instead (see Speaker labels).
Supported Audio Formats
Available Models
Use larger models for:
- Noisy audio
- Accented speech
- Technical vocabulary
Output Formats
Text
Plain text transcript:JSON
Verbose JSON
Includes language, duration, processing-time, realtime-factor, and optional runtime diagnostics. Addtimestamp_granularities[]=word (CLI
--word-timestamps) to get a words array.
Tips for Best Results
- Use quality audio — Clear recordings transcribe better
- Minimize noise — Background noise reduces accuracy
- Proper format — WAV files work best
- Right model size — Larger models for difficult audio
- Language hints — Specify language if known
See Also
- Diarization — Identify multiple speakers
- Speaker Attributed ASR — Speaker-turn transcripts
- Voice Mode — Real-time transcription
- CLI Reference — Full command documentation