Skip to content

Open-source speech recognition toolkit for training, inference, streaming ASR, VAD, punctuation, speaker diarization pipelines, and OpenAI-compatible/MCP serving.

fix(websocket): pass client hotwords to AutoModel as words

Tanishq7d6b06310453
diff --git a/runtime/python/websocket/funasr_wss_server.py b/runtime/python/websocket/funasr_wss_server.py
index 2a449e642..738317746 100644
--- a/runtime/python/websocket/funasr_wss_server.py
+++ b/runtime/python/websocket/funasr_wss_server.py
@@ -155,6 +155,25 @@ def _safe_int(v, default):
return default
+def _model_hotwords(hotwords):
+ """Turn protocol hotwords into the space-separated string AutoModel reads.
+
+ Clients send hotwords as a JSON object string with weights, such as
+ '{"阿里巴巴": 20, "通义实验室": 30}' (runtime/docs/websocket_protocol.md).
+ AutoModel splits a hotword string on whitespace and takes no weights, so
+ keep only the words, as the C++ server does for NN hotwords. Any other
+ value is returned unchanged.
+ """
+ if isinstance(hotwords, str):
+ try:
+ parsed = json.loads(hotwords)
+ except ValueError:
+ return hotwords
+ if isinstance(parsed, dict):
+ return " ".join(parsed)
+ return hotwords
+
+
# ========= speaker db:加缓存,避免每段都读盘 =========
_SPEAKER_DB_CACHE = {}
_SPEAKER_DB_CACHE_TS = 0.0
@@ -502,7 +521,7 @@ async def ws_serve(websocket, path=None):
]
if "hotwords" in messagejson:
- hotword_data = messagejson["hotwords"]
+ hotword_data = _model_hotwords(messagejson["hotwords"])
websocket.status_dict_asr["hotword"] = hotword_data
websocket.status_dict_asr_online["hotword"] = hotword_data
print(f"热词已更新: {hotword_data}")
diff --git a/tests/test_websocket_server_hotwords.py b/tests/test_websocket_server_hotwords.py
new file mode 100644
index 000000000..36b0daeb3
--- /dev/null
+++ b/tests/test_websocket_server_hotwords.py
@@ -0,0 +1,121 @@
+"""Client hotwords in runtime/python/websocket/funasr_wss_server.py.
+
+The websocket protocol (runtime/docs/websocket_protocol.md) sends ``hotwords``
+as a JSON object string that maps each hotword to a weight, for example
+'{"阿里巴巴":20,"通义实验室":30}'. The HTML5, Java, Android and C++ clients
+and the microphone mode of funasr_wss_client.py all send this format. The
+Python server passed the string to AutoModel unchanged, and AutoModel splits a
+hotword string on whitespace, so the JSON text became ``<unk>`` hotwords.
+"""
+
+import asyncio
+import importlib.util
+import json
+import sys
+import types
+from pathlib import Path
+from types import SimpleNamespace
+
+import pytest
+
+ROOT = Path(__file__).resolve().parents[1]
+SERVER_PATH = ROOT / "runtime/python/websocket/funasr_wss_server.py"
+
+
+class FakeWebSocket:
+ def __init__(self, incoming):
+ self.incoming = iter(incoming)
+ self.sent = []
+ self.path = None
+
+ def __aiter__(self):
+ return self
+
+ async def __anext__(self):
+ try:
+ return next(self.incoming)
+ except StopIteration:
+ raise StopAsyncIteration
+
+ async def send(self, message):
+ self.sent.append(message)
+
+ async def close(self):
+ pass
+
+
+def server(monkeypatch):
+ for name in ("websockets", "scipy", "torch"):
+ pytest.importorskip(name)
+ # The server builds its models with AutoModel when it is imported.
+ fake_funasr = types.ModuleType("funasr")
+ fake_funasr.AutoModel = lambda **kwargs: SimpleNamespace(**kwargs)
+ monkeypatch.setitem(sys.modules, "funasr", fake_funasr)
+ monkeypatch.setattr(sys, "argv", [str(SERVER_PATH)])
+ spec = importlib.util.spec_from_file_location("funasr_wss_server_hotwords", SERVER_PATH)
+ module = importlib.util.module_from_spec(spec)
+ spec.loader.exec_module(module)
+ yield module
+ module.EXECUTOR.shutdown(wait=True)
+
+
+def _offline_hotwords(server, hotwords):
+ """Send one offline request and return the hotword AutoModel.generate receives."""
+ asr_calls = []
+
+ def asr_generate(input, **kwargs):
+ asr_calls.append(kwargs)
+ return [{"text": "阿里巴巴"}]
+
+ def no_speaker_model(*args, **kwargs):
+ raise RuntimeError("no speaker model in this test")
+
+ server.model_asr = SimpleNamespace(generate=asr_generate)
+ server.model_vad = SimpleNamespace(generate=lambda input, **kwargs: [{"value": []}])
+ server.model_sv = SimpleNamespace(generate=no_speaker_model)
+ server.model_punc = None
+ config = {
+ "mode": "offline",
+ "chunk_size": [5, 10, 5],
+ "chunk_interval": 10,
+ "wav_name": "demo",
+ "is_speaking": True,
+ "hotwords": hotwords,
+ "itn": True,
+ }
+ websocket = FakeWebSocket(
+ [
+ json.dumps(config, ensure_ascii=False),
+ b"\x01\x00" * 1600,
+ json.dumps({"is_speaking": False, "is_end": True}),
+ ]
+ )
+
+ asyncio.run(server.ws_serve(websocket))
+
+ assert json.loads(websocket.sent[-1])["is_final"] is True
+ assert len(asr_calls) == 1
+ return asr_calls[0]["hotword"]
+
+
+ "hotwords, expected",
+ [
+ # HTML5, Java, Android and C++ clients (compact JSON)
+ ('{"阿里巴巴":20,"通义实验室":30}', "阿里巴巴 通义实验室"),
+ # funasr_wss_client.py microphone mode (json.dumps)
+ ('{"阿里巴巴": 20, "通义实验室": 30}', "阿里巴巴 通义实验室"),
+ # an empty object: no hotwords
+ ("{}", ""),
+ ],
+)
+def test_protocol_hotwords_reach_the_model_as_words(server, hotwords, expected):
+ assert _offline_hotwords(server, hotwords) == expected
+
+
+# A space-separated string is what AutoModel expects already, and
+# funasr_wss_client.py sends "" when --hotword is not set.
[email protected]("hotwords", ["阿里巴巴 通义实验室", ""])
+def test_plain_hotword_strings_are_unchanged(server, hotwords):
+ assert _offline_hotwords(server, hotwords) == hotwords