# OWSM-CTC

OWSM-CTC is ESPnet's non-autoregressive multilingual speech recognition and
speech translation family. The v4 checkpoint is
[espnet/owsm_ctc_v4_1B](https://huggingface.co/espnet/owsm_ctc_v4_1B).
It uses an E-Branchformer encoder with self-conditioned CTC and is separate
from the autoregressive [OWSM](owsm.md) family.

## Usage

```bash
audiocpp_cli --task asr --family owsm_ctc \
  --model /path/to/OWSM-CTC-GGUF/owsm-ctc-v4-1b-f32.gguf \
  --backend cuda --audio input.wav \
  --request-option language=eng --text-out transcript.txt --log
```

Add `--request-option target_language=deu` to translate into German instead
of transcribing. Language values are the checkpoint's ISO 639-3 codes, such as
`eng`, `jpn`, and `zho`. `language=auto` uses unknown-language conditioning;
it does not report a detected language.

The model runs offline best-path CTC decoding. It does not generate timestamp
tokens, use an autoregressive beam, or condition on previous transcript text.

## Common Options

| CLI option | Description |
| --- | --- |
| `--audio` | Input audio; resampled to mono 16 kHz. |
| `--backend` | Execution backend. |
| `--mode` | `offline` (default) or `streaming` text output from a complete audio file. |
| `--threads` | CPU thread count. |
| `--text-out` | Transcript output path. |

## Request Options (Use With `--request-option`)

| Option | Default | Description |
| --- | --- | --- |
| `language` | `eng` | Input language code; `auto` uses unknown-language conditioning. |
| `target_language` | Unset | Translation target language; omit for transcription. |
| `audio_chunk_mode` | `auto` | `auto`, `fixed`, or `none`. |

## Session Options (Use With `--session-option`)

| Option | Default | Description |
| --- | --- | --- |
| `owsm_ctc.weight_type` | `native` | Matmul weight storage override. |

## Long Audio

`auto` pads short input to the trained 30-second window. Longer recordings use
overlapping 30-second windows with two seconds of context on each side.
CTC frame predictions are joined before collapsing repeated tokens and blanks.
`fixed` also applies this overlapping-window path to short audio. `none` rejects
input longer than 30 seconds. The result is a whole-recording transcript,
without word or segment timestamps.

## Streaming

Add `--mode streaming` for incremental text output from a complete audio file.
Each completed long-form window emits only newly committed text. CTC repeat
and blank collapsing continues across window boundaries. Short input produces
one text update after its window completes. The final transcript is identical
to offline mode with the same settings. This is not live microphone streaming.

For the server, configure the model with `"mode": "streaming"` and send:

```bash
curl http://localhost:8080/v1/audio/transcriptions \
  -F model=owsm_ctc -F file=@input.wav -F language=eng -F stream=true
```

The response uses `transcript.text.delta` followed by `transcript.text.done`.

## Conversion

```bash
python tests/owsm_ctc/convert_gguf.py /path/to/owsm_ctc_v4_1B \
  /path/to/OWSM-CTC-GGUF/owsm-ctc-v4-1b-f32.gguf \
  --staging-dir /path/to/staging --type orig
```

Use `--type q8_0` and a corresponding output filename for Q8. The GGUF includes
the tokenizer, configuration, normalization statistics, and model spec.