---
language:
- zh
base_model:
- JacobLinCool/whisper-large-v3-turbo-common_voice_19_0-zh-TW
pipeline_tag: automatic-speech-recognition
tags:
  - audio
  - automatic-speech-recognition
library_name: ctranslate2
---

# asadfgglie/faster-whisper-large-v3-zh-TW

此模型是將[JacobLinCool/whisper-large-v3-turbo-common_voice_19_0-zh-TW](https://huggingface.co/JacobLinCool/whisper-large-v3-turbo-common_voice_19_0-zh-TW)
轉換成`CTranslate2`格式的模型，可以在[faster-whisper](https://github.com/systran/faster-whisper)中使用。

## Example

```python
from faster_whisper import WhisperModel

model = WhisperModel("asadfgglie/faster-whisper-large-v3-zh-TW")

segments, info = model.transcribe("audio.mp3")
for segment in segments:
    print("[%.2fs -> %.2fs] %s" % (segment.start, segment.end, segment.text))
```

## Conversion details

原始模型是根據以下指令轉換:

```
ct2-transformers-converter --output_dir faster-whisper-large-v3-zh-TW \
 --model JacobLinCool/whisper-large-v3-turbo-common_voice_19_0-zh-TW \
 --copy_files preprocessor_config.json
```

在轉換完成後，請記得自行到原始模型的[model card](https://huggingface.co/openai/whisper-large-v3)中下載`tokenizer.json`。
(因為`JacobLinCool/whisper-large-v3-turbo-common_voice_19_0-zh-TW`的repo中沒有，而`faster_whishper`又需要這個酷東東來做tokenizer)

如果有需要，你可以在轉換指令中添加`--quantization float16`來指定量化精度。不過在推理時你依舊可以使用[`compute_type`](https://opennmt.net/CTranslate2/quantization.html)參數來進一步量化/去除量化。