feat: 启用豆包二遍识别模式以提升实时性和准确率

- 切换到 bigmodel_async endpoint 并启用 enable_nonstream
- 第一遍流式识别提供实时文字预览
- VAD 分句后自动触发第二遍非流式识别提升准确率
- 修改文本处理逻辑从累加改为替换(适配 full 模式)
- 统一配置字段命名:app_key → app_id, access_key → access_token
This commit is contained in:
2026-03-01 21:34:54 +08:00
parent e4b5841c93
commit 8c7b9b45fd
7 changed files with 55 additions and 53 deletions

View File

@@ -104,13 +104,14 @@ type AudioMeta struct {
}
type RequestMeta struct {
ModelName string `json:"model_name"`
EnableITN bool `json:"enable_itn"`
EnablePUNC bool `json:"enable_punc"`
EnableDDC bool `json:"enable_ddc"`
ShowUtterances bool `json:"show_utterances"`
ResultType string `json:"result_type,omitempty"`
EndWindowSize int `json:"end_window_size,omitempty"`
ModelName string `json:"model_name"`
EnableITN bool `json:"enable_itn"`
EnablePUNC bool `json:"enable_punc"`
EnableDDC bool `json:"enable_ddc"`
ShowUtterances bool `json:"show_utterances"`
ResultType string `json:"result_type,omitempty"`
EnableNonstream bool `json:"enable_nonstream,omitempty"`
EndWindowSize int `json:"end_window_size,omitempty"`
}
// EncodeFullClientRequest builds the binary message for the initial handshake.
// nostream mode: header(4) + payload_size(4) + gzip(json)