Bold words are the pair of words selected from the rhyming buckets: a/aA and u/U kya hai mere liye jindagiДЇigure 3: Example of A BA B rhyme scheme (LSTM model). jo haath kaam ud khul main sharamaati phiruДЇigure 2: Example of A BBA rhyme scheme (LSTM model). duniya hu dil se tu mujhe kya kho jayega yun aku ta ruswaii mujhako sharmaata tha beete saaye se hone laga hoon kabhi sochataa hoon main khel kahu ya nindiya Figures 3 and 2 are the examples of the output generated from Bollyrics for ABAB and ABBA rhyme schemes respectively. We repeat the steps for all the paragraphs. Then, we reverse the generated sentenceвЂ™s ordering to obtain a line of the lyrics. Using the outputs and a time step of two words, we sequentially generate a sentence until вЂ™ вЂ™ token appears or the sentence length becomes more than 10. We reverse the order of the words and inference the model. Then, we start iterating through each letter in the scheme and select a pair of words randomly from the group corresponding to that letter. The number of letters in the scheme denotes the number of lines in each paragraph, and the number of paragraphs in the generated song is decided by the user. We randomly allocate a unique group from Table 2 to each letter in the scheme without replacement. Rhyme scheme can be of any form - AABB/ABAB/AAA/ABCD and many more. To generate lyrics through Bollyrics, the user can provide a rhyme scheme for the paragraph. Finally, we automatically mapped the ending character with its Devanagri sound to get 13 groups of sounds shown in Table 2. The above process reduces the vocabulary to 70,731 from 71,153.

We performed this process on top 1000 most-occurring ending words based on our understanding of Hindi langauge. Since our dataset is in the romanized script, we grouped the words by manually mapping a wordвЂ™s ending character with itвЂ™s corresponding sound in Devanagri script to account for variation. E.g., the Hindi word for вЂњweвЂќ can be written as вЂњhumeinвЂќ, вЂњhumeвЂќ, and вЂњhumeyвЂќ in romanized version. Though it is trivial to group together words in the original script of a language, the machine-translated version of a language may result in varying ending letters. Hence, we group the words based on the last syllable in our corpus. For e.g., вЂњColonelвЂќ is pronounced as вЂњk ernlвЂќ by skipping the character вЂњlвЂќ and adding new sounds in English, whereas in Hindi вЂњColonelвЂќ is written as вЂњkarnalвЂќ, whose characterвЂ™s sounds aligns with itвЂ™s pronunciation. As opposed to English, all the characters sound corresponds to a word pronunciation in Hindi. To generate lyrics with rhyming schemes, we need to apply a grouping technique to collect words with similar ending sounds. ( 2018)) we have introduced an unique approach to account for rhyming patterns in Hindi language. ( 2015), Peterson ( 2019)) and Chinese languages ( Wu et al. Since majority of the work is conducted in English ( Potash et al. Such properties distinguish normal text from lyrics/songs. Apart from utilizing the power of text generation models to capture context, Bollyrics capture the song structure (i.e., account for the number of paragraphs and lines), and rhythmic patterns (E.g., AABB, ABBA, ABCD) to generate a particular song. Hence, we approach our problem as a text generation task.

These models have shown promising results in generating human-readable and semantically significant creative content. To generate human-like texts by training billions of parameters and text documents. Researchers have developed highly advanced language models such as BERT Devlin et al. With the advent of deep learning in natural language processing, machines learn and predict the context with high precision.

Lyrics generator