Open-source speech recognition toolkit for training, inference, streaming ASR, VAD, punctuation, speaker diarization pipelines, and OpenAI-compatible/MCP serving.
Python asraudiochineseemotion-recognitionfunasrmcp-servermultilingual-asropenai-compatible-apiparaformerpunctuationpytorchreal-time-asrspeaker-diarizationspeech-recognitionspeech-to-textstreaming-asrtranscriptionvllmvoice-activity-detectionwhisper-alternative
"""Regression tests for empty text in the realtime CT-Transformer punctuation model."""
import unittest
import torch
from funasr.auto.auto_model import AutoModel
from funasr.models.ct_transformer_streaming.model import CTTransformerStreaming
from funasr.tokenizer.char_tokenizer import CharTokenizer
PUNC_LIST = ["<unk>", "_", ",", "。", "?", "、"]
TOKEN_LIST = ["<blank>", "<s>", "</s>", "<unk>", "跨", "境", "河", "流", "是"]
def _tiny_streaming_punc_model():
"""Build a small randomly initialised CTTransformerStreaming that never adds punctuation."""
model = CTTransformerStreaming(
encoder="SANMVadEncoder",
encoder_conf={
"input_size": 8,
"output_size": 8,
"attention_heads": 2,
"linear_units": 16,
"num_blocks": 2,
"dropout_rate": 0.0,
"positional_dropout_rate": 0.0,
"attention_dropout_rate": 0.0,
"input_layer": "pe",
"kernel_size": 3,
"sanm_shfit": 0,
"padding_idx": 0,
},
vocab_size=len(TOKEN_LIST),
punc_list=PUNC_LIST,
embed_unit=8,
att_unit=8,
sentence_end_id=3,
)
with torch.no_grad():
model.decoder.weight.zero_()
model.decoder.bias.zero_()
model.decoder.bias[PUNC_LIST.index("_")] = 1.0
return model.eval()
class TestCTTransformerStreamingEmptyInput(unittest.TestCase):
def setUp(self):
self.model = _tiny_streaming_punc_model()
self.tokenizer = CharTokenizer(token_list=TOKEN_LIST, unk_symbol="<unk>")
def test_empty_text_returns_empty_result(self):
for text in ("", " "):
with self.subTest(text=text):
results, _ = self.model.inference(
[text], key=["utt"], tokenizer=self.tokenizer, cache={}, device="cpu"
)
self.assertEqual(results, [{"key": "utt", "text": "", "punc_array": None}])
def test_empty_segment_keeps_pending_words(self):
am = AutoModel.__new__(AutoModel)
am.model = self.model
am.vad_model = None
am.punc_model = None
am.kwargs = {"device": "cpu", "disable_pbar": True, "tokenizer": self.tokenizer}
cache = {}
self.assertEqual(am.generate(input="跨境河流", cache=cache)[0]["text"], "跨境河流")
self.assertEqual(cache["pre_text"], ["跨", "境", "河", "流"])
self.assertEqual(am.generate(input="", cache=cache)[0]["text"], "")
self.assertEqual(cache["pre_text"], ["跨", "境", "河", "流"])
self.assertEqual(am.generate(input="是", cache=cache)[0]["text"], "是")
if __name__ == "__main__":
unittest.main()