Whisper Large-v3 Turbo
此 Soniqo 页面记录本地 speech-swift / speech-core 实现中的 Whisper Large-v3 Turbo。Hugging Face 包链接放在集成说明之后。
先进入站内页面
首页卡片和文档菜单先指向这里；源模型和权重包链接仍在本页提供。
概览
| 模型 | Whisper Large-v3 Turbo |
|---|---|
| 用途 | General multilingual speech-to-text |
| 后端 | CoreML fp16 on CPU and Neural Engine |
| 输出 | Text transcript |
| 语言 | Whisper multilingual set, about 100 languages |
| 许可证 | MIT weights from OpenAI |
| 状态 | Ready through speech transcribe --engine whisper and the WhisperASR Swift product |
| 来源 | openai/whisper-large-v3-turbo |
| Swift 产品 | WhisperASR |
| CLI / 运行时 | speech transcribe --engine whisper |
使用
下面的片段对应当前 speech-swift 仓库暴露的 API 或命令。
# Transcribe with the native CoreML Whisper runtime.
.build/release/speech transcribe recording.wav --engine whisper
.build/release/speech transcribe recording.wav --engine whisper --language en模型链接
实现说明
- LibriSpeech test-clean slice on an M5 Pro: 1.40% WER, mean RTF 0.089, 6.1 s model load, 384 MB peak RSS — versus 1.53% / 0.085 / 100.2 s / 507 MB for a direct WhisperKit run of the same model.
- The bundle splits into four CoreML models — mel spectrogram, audio encoder, decoder context prefill, and a KV-cached text decoder — with the encoder and decoder on the Neural Engine.
- Greedy no-timestamp decoding with automatic language detection or a --language hint; audio is processed in 30 s chunks with a repeated-word guard against greedy hallucination loops.
- Word timestamps, temperature fallback, and VAD-guided long-form seeking are not implemented yet; --model accepts turbo, large-v3-turbo, or any Hugging Face repo ID.