[tanishqbaweja]

Summary

Fixes #3886

Since #2791, runtime/python/websocket/funasr_wss_server.py built its offline model from a fixed "paraformer-zh" and "v2.0.4", although it parses --asr_model and --asr_model_revision and the README lists --asr_model (runtime/python/websocket/README.md:29). Both options had no effect: a server started with --asr_model paraformer-en, or with the directory of a fine-tuned model, still ran paraformer-zh for its offline and 2pass-offline results.

This PR builds the offline model from the two options again, as the streaming, VAD and punctuation models are built from theirs (two lines). So that a server started without these options keeps its model, the default of --asr_model becomes "paraformer-zh" (one line). It named the Contextual Paraformer (iic/speech_paraformer-large-contextual_asr_nat-zh-cn-16k-common-vocab8404), which the server did not load. --asr_model_revision already defaults to v2.0.4.

The model and revision that the offline model's AutoModel call receives (the script in the issue):

options	main	this PR
none	paraformer-zh v2.0.4	unchanged
--asr_model paraformer-en	paraformer-zh v2.0.4	paraformer-en v2.0.4
--asr_model /models/my-paraformer --asr_model_revision master	paraformer-zh v2.0.4	/models/my-paraformer master

Type of change

• Bug fix
• Documentation
• Example or demo
• Runtime or deployment
• Benchmark or evaluation
• Model/training change

Validation

The new tests/test_websocket_server_asr_model.py runs the real funasr_wss_server.py module (the server builds its models when it is imported) with AutoModel replaced by a recorder, and with empty stand-ins for websockets, scipy and torch, which the server uses only to serve requests. So it needs only numpy. It checks the model and revision of the offline model without options, with --asr_model paraformer-en, and with a model directory and --asr_model_revision master, and that the streaming model keeps --asr_model_online.

Windows 11, CPU, Python 3.11.17, numpy 1.26.4, pytest 9.2.0:

python -m pytest -q tests/test_websocket_server_asr_model.py tests/test_websocket_file_finalization.py

• main (6f52414) server with the new test: 2 failed, 29 passed: assert 'paraformer-zh' == 'paraformer-en' and assert 'paraformer-zh' == '/models/my-paraformer'.

• This branch: 31 passed, also with -W error.

• python -m compileall funasr examples tests (and runtime/python/websocket)

• Docs or links checked

• Runtime/deployment command tested (the server's startup: its real argument parsing and model construction, with AutoModel replaced; no model was loaded and no socket was served)

black --line-length=100 (24.4.0, as in pre-commit) leaves the new test unchanged, and makes the same changes to funasr_wss_server.py on main and on this branch, none on the changed lines.

User impact

Users who start the Python websocket server with --asr_model, and optionally --asr_model_revision, get that model for offline and 2pass-offline recognition, for example an English model or a fine-tuned model directory. Before, the server ran paraformer-zh v2.0.4 whatever they passed. A server started without these options loads paraformer-zh v2.0.4 as before.

Notes for reviewers

• The new default of --asr_model is the model the server has loaded since Codex/create training and inference scripts #2791, not the Contextual Paraformer that the option named, so that servers started without options keep their model.
• No workflow's path filters cover runtime/python/websocket/**, so CI does not run this test (or tests/test_websocket_file_finalization.py). I can add it to a workflow if you want.
• git merge-tree --write-tree of this branch with the heads of fix(websocket): pass client hotwords to AutoModel as words #3880 and Fix/security hardening #3877, which also edit funasr_wss_server.py, gives no conflicts. On the tree merged with fix(websocket): pass client hotwords to AutoModel as words #3880, this test, its tests/test_websocket_server_hotwords.py and tests/test_websocket_file_finalization.py give 36 passed (run with stand-in websockets, scipy and torch modules, which my test environment does not have).
• No version bump and no workflow change.