Installation

We adapted this ctc decoder from here. This decoder can only run on cpu.

• continuous decoding for streaming asr
• support kenlm language model
• multiprocessing

To install the decoder:

git clone https://github.com/Slyne/ctc_decoder.git
apt-get update
apt-get install swig
apt-get install python3-dev 
cd ctc_decoder/swig && bash setup.sh

Usage

Please refer to swig/test/test_en.py and swig/test/test_zh.py for how to do streaming decoding and offline decoding w/o language model.

Adding language model

How to build the language model ? You may refer to kenlm. For Mandarin, the input text for language model should be like:

好 好 学 习 ，天 天 向 上 ！
再 接 再 厉
...

There's a space between two characters.

For English, the input text is just like the normal text.

Share Market Today - Stock Market and Share Market Live Updates

How to add language model:

alpha = 0.5
beta = 0.5
lm_path = '../kenlm/lm/test.arpa'
scorer = decoder.Scorer(alpha, beta, lm_path, vocab_list)
......
result1 =  decoder.ctc_beam_search_decoder_batch(batch_chunk_log_prob_seq, 
                                                 batch_chunk_log_probs_idx,
                                                 batch_root_trie,
                                                 batch_start,
                                                 beam_size, num_processes,
                                                 blank_id, space_id,
                                                 cutoff_prob, scorer)

How language model in called in this implementation of ctc prefix beam search ?

If the language model is char based (like the Mandarin lm), it will call the language model scorer all the times. If the language model is word based (like the English lm), it will only call the scorer whenever space_id is detected.

Adding hotwords

Please refer to the following steps how to use hotwordsboosting.

• Step 1. Initialize HotWordsScorer

# if you don't want to use hotwords. set hotwords_scorer=None(default), 
# vocab_list is Chinese characters.
hot_words = {'再接': 10, '再厉': -10, '好好学习': 100}
hotwords_scorer = HotWordsScorer(hot_words, vocab_list, is_character_based=True)

If you set is_character_based is True (default mode), the first step is to combine Chinese characters into words, if words in hotwords dictionary then add hotwords score. If you set is_character_based is False, all words in the fixed window will be enumerated.

• Step 2. Add hotwords_scorer when decoding

result2 =  decoder.ctc_beam_search_decoder_batch(batch_chunk_log_prob_seq, 
                                                 batch_chunk_log_probs_idx,
                                                 batch_root_trie,
                                                 batch_start,
                                                 beam_size, num_processes,
                                                 blank_id, space_id,
                                                 cutoff_prob, scorer, hotwords_scorer)

Please refer to swig/test/test_zh.py for how to decode with hotwordsboosting.