见独技术

纯识别(/v1/asr)

实时送入中文普通话音频,识别过程中返回中间结果,每段话结束返回定稿文本与起止时间;会话结束后返回全部分段。

入口 wss://api.jftxyun.com:31080/v1/asr,握手头 Authorization: Bearer <纯识别 key>。key、错误码、费用、查询接口见接入说明。

会话流程

-> session.start        {"type":"session.start","session_id":"…","features":{"max_call_seconds":600}}
<- session.ready        {"input_audio":{"encoding":"pcm_s16le","sample_rate":16000,…}}
-> input_audio.append   {"session_id":"…","sequence":1,"audio":"<16 kHz PCM16 单声道 base64>"}   sequence 从 1 连续
<- speech.started       {"turn_id":"turn_000001"}                    一段话开始
<- asr.partial          {"turn_id":…,"sequence":n,"text":"本段累计文本","text_mode":"cumulative"}
<- asr.final            {"turn_id":…,"text":"…","audio_start_ms":…,"audio_end_ms":…}
-> session.close
<- session.result       {"status":"processing"}
<- session.result       {"status":"completed","turns":[{turn_id,text,audio_start_ms,audio_end_ms}],"audio_ms":…,"billing":{…}}
-> session.result.ack   {"result_event_id":"<completed 那条 session.result 的 event_id>"}
<- session.closed       {"reason":"completed"|"ack_timeout"}

说明

示例代码(Python)

读 wav,按实时节奏推流,打印中间结果、定稿与会后结果。依赖 Python 3.10+ 与 websockets>=14;wav 要求单声道 16 bit,采样率不是 16 kHz 时脚本会重采样(电话 8 kHz 录音可直接用)。

pip install "websockets>=14"
export PAA_API_KEY=sk_live_...          # 纯识别的 key
python3 asr_api_sample.py a.wav [b.wav ...]
加载中…