Instructions to use espnet/owsm_v4_base_102M with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- ESPnet
How to use espnet/owsm_v4_base_102M with ESPnet:
import soundfile from espnet2.bin.asr_inference import Speech2Text model = Speech2Text.from_pretrained( "espnet/owsm_v4_base_102M" ) speech, rate = soundfile.read("speech.wav") text, *_ = model(speech)[0] - Notebooks
- Google Colab
- Kaggle
Link the community audio.cpp GGUF builds
#2
by sw005320 - opened
README.md
CHANGED
|
@@ -39,6 +39,8 @@ The code, pre-trained model weights, and training logs are publicly released to
|
|
| 39 |
Inference examples can be found on our [project page](https://www.wavlab.org/activities/2024/owsm/).
|
| 40 |
The Gradio demo is [here](https://huggingface.co/spaces/espnet/owsm-v4), which runs the 1B model.
|
| 41 |
|
|
|
|
|
|
|
| 42 |
[OWSM v4](https://www.isca-archive.org/interspeech_2025/peng25c_interspeech.html) is the latest version in the OWSM series, which significantly outperforms OWSM v3.1 in LID and multilingual ASR.
|
| 43 |
Additionally, OWSM v4 applies 8 times subsampling (instead of 4 times in OWSM v3.1) to the log Mel features, leading to a final resolution of 80 ms in the encoder.
|
| 44 |
When running inference, we recommend setting `maxlenratio=1.0` (default) instead of smaller values.
|
|
|
|
| 39 |
Inference examples can be found on our [project page](https://www.wavlab.org/activities/2024/owsm/).
|
| 40 |
The Gradio demo is [here](https://huggingface.co/spaces/espnet/owsm-v4), which runs the 1B model.
|
| 41 |
|
| 42 |
+
For CPU inference without Python, GGUF conversions of this model are available at [audio-cpp/audio.cpp-gguf](https://huggingface.co/audio-cpp/audio.cpp-gguf/tree/main/OWSM-GGUF) (`owsm-v4-base-102m`), for use with [audio.cpp](https://github.com/0xShug0/audio.cpp). These builds are produced and maintained outside ESPnet.
|
| 43 |
+
|
| 44 |
[OWSM v4](https://www.isca-archive.org/interspeech_2025/peng25c_interspeech.html) is the latest version in the OWSM series, which significantly outperforms OWSM v3.1 in LID and multilingual ASR.
|
| 45 |
Additionally, OWSM v4 applies 8 times subsampling (instead of 4 times in OWSM v3.1) to the log Mel features, leading to a final resolution of 80 ms in the encoder.
|
| 46 |
When running inference, we recommend setting `maxlenratio=1.0` (default) instead of smaller values.
|