diff --git a/test/media.test.js b/test/media.test.js index 22e3a14..01aaa5f 100644 --- a/test/media.test.js +++ b/test/media.test.js @@ -18,6 +18,7 @@ import { import { createStreamingWebmDecoder, downsamplePcm16, + findEbmlHeader, sniffAudioFormat, webmToPcm, } from "../agent/infra/media/audio.js"; @@ -51,6 +52,15 @@ test("按文件头识别音频容器,开发者工具 WebM 不会被当成 mp3" assert.equal(sniffAudioFormat(Buffer.alloc(0)), "unknown"); }); +test("首片带前导字节时能在头部窗口内定位 EBML 头", () => { + const magic = Buffer.from([0x1a, 0x45, 0xdf, 0xa3]); + const padded = Buffer.concat([Buffer.from([0, 0, 0]), magic, Buffer.alloc(8)]); + assert.equal(findEbmlHeader(padded), 3); + assert.equal(findEbmlHeader(Buffer.concat([magic, Buffer.alloc(8)]), 4096), 0); + // 纯编码数据里不该出现魔数,扫描结果必须是 -1 + assert.equal(findEbmlHeader(Buffer.from([0xff, 0xf3, 0x84, 0xc0, 0x00, 0x11]), 4096), -1); +}); + test("WebM 解码结果下混为单声道 16bit PCM", async () => { const decodeImpl = async () => ({ channelData: [new Float32Array([0.5, -0.5, 1.5, -1.5])], @@ -202,6 +212,80 @@ test("DashScope 会话缺密钥时在本地失败且不建连", async () => { await assert.rejects(() => session.finish(), /api key missing/); }); +test("DashScope 实时会话把中间结果经 onPartial 回传,末帧定稿", async () => { + /** 假 WebSocket:只回放本用例关心的事件 */ + class FakeWs { + constructor(url, options) { + this.url = url; + this.options = options; + this.readyState = 1; + this.bufferedAmount = 0; + this.sent = []; + this.listeners = {}; + FakeWs.instances.push(this); + } + on(event, callback) { + this.listeners[event] = callback; + } + emit(event, data) { + this.listeners[event]?.(data); + } + send(data) { + this.sent.push(data); + } + close() { + this.readyState = 3; + } + } + FakeWs.instances = []; + + const partials = []; + const session = createRecognizer({ + provider: "dashscope", + endpoint: "wss://example.invalid/asr", + apiKey: "test-key", + logger: silentLogger, + // 关掉节流,断言每一个中间结果都回调 + partialIntervalMs: 0, + WebSocketImpl: FakeWs, + onPartial: (text) => partials.push(text), + }); + + session.push(Buffer.from([1, 2, 3, 4])); + const ws = FakeWs.instances[0]; + assert.ok(ws, "首帧应触发建连"); + + // 建连后先发 run-task,再在 task-started 时把缓存的音频冲出去 + ws.emit("open"); + const runTask = JSON.parse(ws.sent[0]); + assert.equal(runTask.header.action, "run-task"); + assert.equal(runTask.header.task_id.length > 0, true); + + ws.emit("message", JSON.stringify({ header: { event: "task-started" }, payload: {} })); + assert.equal(Buffer.isBuffer(ws.sent[1]), true); + + // sentence_end=false 是中间结果:必须实时回显(不再等整句说完) + ws.emit("message", JSON.stringify({ + header: { event: "result-generated" }, + payload: { output: { sentence: { text: "你好", sentence_end: false } } }, + })); + assert.deepEqual(partials, ["你好"]); + + ws.emit("message", JSON.stringify({ + header: { event: "result-generated" }, + payload: { output: { sentence: { text: "你好,我是小橙", sentence_end: false } } }, + })); + assert.deepEqual(partials, ["你好", "你好,我是小橙"]); + + // 定稿:句子结束 + finish-task 后的 task-finished 取全量 + const finished = session.finish(); + ws.emit("message", JSON.stringify({ + header: { event: "task-finished" }, + payload: { output: { text: "你好,我是小橙。" } }, + })); + assert.match(await finished, /你好,我是小橙。/); +}); + test("批式 OCR 会话累积分片并传递元信息", async () => { let captured = null; const session = createOcrSession({