o }oi|b@sddlZddlZddlZddlmZddlmZmZmZm Z m Z m Z m Z ddl mZddlmZmZmZmZddlmZddlmZmZmZddlmZdd lmZeGd d d eZdS) N) defaultdict)CallableDictListOptionalSetTupleUnion)validate_locale)LATIN_CHARS_ALLany_locale_word_tokenizeenglish_word_tokenizenormalize_unicode_text)BaseG2p)GRAPHEME_CASE_MIXEDGRAPHEME_CASE_UPPERset_grapheme_case)logging) experimentalcseZdZddgZededZededZddddd dde d df d e e e j ee eee ffd e d eee ge fdedee e e j ee fdedeedeedee dee dee ddffdd Zed e e e j ee eee ffdee eee ffddZd e e e j ee eee fffddZede e e j fdee fddZd e dee fd!d"Zd#ee eee fdeee eee feffd$d%Zd/d&d'Zd e defd(d)Zd e deee effd*d+Zd,e dee fd-d.Z Z!S)0IpaG2puˈuˌ[z\d]z[^en-USNTF phoneme_dictlocaleapply_to_oov_wordignore_ambiguous_words heteronyms use_charsphoneme_probability use_stresses grapheme_casegrapheme_prefix mapping_filereturnc sD|_| _| _|_|_t_|durt||s,jdur,d_ t dn|_ |} | r? | \} _nt|d|durOt d|dkrVt}nt}tj| ||| d|_t|tspt|tjryt|_nt|trtdd |Drt|_nd_jrfd d jD_dSdS) u Generic IPA G2P module. This module converts words from graphemes to International Phonetic Alphabet representations. Optionally, it can ignore heteronyms, ambiguous words, or words marked as unchangeable by `word_tokenize_func` (see code for details). Ignored words are left unchanged or passed through `apply_to_oov_word` for handling. Args: phoneme_dict (str, Path, or Dict): Path to file in CMUdict format or an IPA dict object with CMUdict-like entries. For example, a dictionary file: scripts/tts_dataset_files/ipa_cmudict-0.7b_nv22.06.txt; a dictionary object: {..., "Wire": [["ˈ", "w", "a", "ɪ", "ɚ"], ["ˈ", "w", "a", "ɪ", "ɹ"]], ...}. locale (str): Locale used to determine a locale-specific tokenization logic. Currently, it supports "en-US", "de-DE", and "es-ES". Defaults to "en-US". Specify None if implementing custom logic for a new locale. apply_to_oov_word (Callable): Function that deals with the out-of-vocabulary (OOV) words that do not exist in the `phoneme_dict`. ignore_ambiguous_words (bool): Whether to handle word via phoneme_dict with ambiguous phoneme sequences. Defaults to True. heteronyms (str, Path, List[str]): Path to file that includes heteronyms (one word entry per line), or a list of words. use_chars (bool): Whether to include chars/graphemes in the token list. It is True if `phoneme_probability` is not None or if `apply_to_oov_word` function ever returns graphemes. phoneme_probability (Optional[float]): The probability (0.0 <= ε <= 1.0) that is used to balance the action that a word in a sentence is whether transliterated into a sequence of phonemes, or kept as a sequence of graphemes. If a random number for a word is greater than ε, then the word is kept as graphemes; otherwise, the word is transliterated as phonemes. Defaults to None which is equivalent to setting it to 1.0, meaning always transliterating the word into phonemes. Note that this code path is only run if the word can be transliterated into phonemes, otherwise, if a word does not have an entry in the g2p dict, it will be kept as graphemes. If a word has multiple pronunciations as shown in the g2p dict and `ignore_ambiguous_words` is True, it will be kept as graphemes as well. use_stresses (Optional[bool]): Whether to include the stress symbols (ˈ and ˌ). grapheme_case (Optional[str]): Trigger converting all graphemes to uppercase, lowercase, or keeping them as original mix-cases. You may want to use this feature to distinguish the grapheme set from the phoneme set if there is an overlap in between. Defaults to `upper` because phoneme set only uses lowercase symbols. You could explicitly prepend `grapheme_prefix` to distinguish them. grapheme_prefix (Optional[str]): Prepend a special symbol to any graphemes in order to distinguish graphemes from phonemes because there may be overlaps between the two set. It is suggested to choose a prefix that is not used or preserved somewhere else. "#" could be a good candidate. Default to "". TODO @borisfom: add docstring for newly added `mapping_file` argument. NTzdphoneme_probability was not None, characters will be enabled even though use_chars was set to False.z contains no entries!aapply_to_oov_word=None, This means that some of words will remain unchanged if they are not handled by any of the rules in self.parse_one_word(). This may be intended if phonemes and chars are both valid inputs, otherwise, you may see unexpected deletions in your input.r)rword_tokenize_funcrr#css|]}t|tVqdS)N) isinstancestr.0hetr+\/home/ubuntu/.local/lib/python3.10/site-packages/nemo/collections/tts/g2p/models/i18n_ipa.py sz"IpaG2p.__init__..csh|] }t|jdqS)case)rr!r(selfr+r, z"IpaG2p.__init__..)r r!r"rrrandomRandom_rngr rrwarning_parse_phoneme_dict_normalize_dictsymbols ValueErrorr r super__init__rr&r'pathlibPathset_parse_file_by_linesrlistall)r1rrrrrrrr r!r"r#phoneme_dict_obj _phoneme_dictr% __class__r0r,r=+sP5    zIpaG2p.__init__cCst|ts t|tjrtt}td}t|ddd}|D]w}t |dkr'qt |}d|dkr7dksrnd|dkrDd ksrnd |dkrQd ksrnd |dkr^d ksrnd|dkrkdksrn|ddkr| j dd}t |d|d}t dd|d}||t|qWd|S1swY|Stdi}|D]%\}}t|tsJdt|dt |}dd|D}|||iq|S)u parse an input IPA dictionary and save it as a dict object. Args: phoneme_dict (Union[str, pathlib.Path, dict]): Path to file in CMUdict format or an IPA dict object with CMUdict-like entries. For example, a dictionary file: scripts/tts_dataset_files/ipa_cmudict-0.7b_nv22.06.txt; a dictionary object: {..., "Wire": [["ˈ", "w", "a", "ɪ", "ɚ"], ["ˈ", "w", "a", "ɪ", "ɹ"]], ...}. Returns: a dict object (Dict[str, List[List[str]]]). z \([0-9]+\)rutf-8encodingrAZazÀÖØöøÿ')maxsplitrz\s+NzGLoading phoneme_dict as a Dict object, and validating its entry format.zPronunciation type is not supported. Please convert to .cSsg|] }dd|DqS)cSsg|]}t|qSr+)r)r)pr+r+r, z9IpaG2p._parse_phoneme_dict...r+)r)pronr+r+r,rZr3z.IpaG2p._parse_phoneme_dict..)r&r'r>r?rrBrecompileopenlenrstripsplitsubappendrinfoitemstypeupdate)rrD_alt_refdictlinepartswordpronsr+r+r,r8sH    ++ zIpaG2p._parse_phoneme_dictcCs|||_dS)zF Replace model's phoneme dictionary with a custom one N)r8r)r1rr+r+r, replace_dictszIpaG2p.replace_dictrYcCsDt|ddd}dd|DWdS1swYdS)NrHrIrJcSsg|]}|qSr+)rstrip)r)rkr+r+r,rZr[z/IpaG2p._parse_file_by_lines..)r_ readlines)rYfr+r+r,rAs$zIpaG2p._parse_file_by_linesrmcsfdd|DS)Ncg|] }j|qSr+r")r) characterr0r+r,rZr3z7IpaG2p._prepend_prefix_for_one_word..r+r1rmr+r0r,_prepend_prefix_for_one_wordz#IpaG2p._prepend_prefix_for_one_wordrDc stt}t}|D]T\}}t|jd}jr(jd|}| |t}j s@|D]} | fdd| Dq0n|}|D]} | | qD|||<jt kr_|s_|||<q ||fS)u Parse a python dict object according to the decision on word cases and removal of lexical stress markers. Args: phoneme_dict_obj (Dict[str, List[List[str]]]): a dictionary object. e.g. {..., "Wire": [["ˈ", "w", "a", "ɪ", "ɚ"], ["ˈ", "w", "a", "ɪ", "ɹ"]], ...} Returns: g2p_dict (dict): processed dict. symbols (set): a IPA phoneme set, or its union with grapheme set. r.rcsg|] }|jvr|qSr+)STRESS_SYMBOLS)r)symbolr0r+r,rZ r3z*IpaG2p._normalize_dict..)rrBr@rfrr!r PUNCT_REGEXrcrhrwr rdrisupperupper) r1rDg2p_dictr:rmrnword_new word_no_punct prons_newr\r+r0r,r9s(   zIpaG2p._normalize_dictcCst|}g}i}|jD]K\}}t|t||j}||} | r(||q g} |D]} t| |} | s;| | q,t| t|krX|sL||q | sT||q | ||<q |D]} |j| =q[|rj|j|||_ dS)a`Replaces the vocabulary of symbols with the one given. Also filters out any entries with illegal graphemes or phonemes according to the new vocab. Args: symbols (List, Set): User-provided set of valid symbols, both graphemes and phonemes keep_alternate (bool): Whether to keep the other pronunciation(s) of a word if not all contain illegal phonemes (and the word doesn't contain illegal graphemes). Warning: this may change a word from being ambiguous to having only one valid pronunciation. Defaults to True. N) r@rrfrwrr!rdr`rhr:)r1r:keep_alternate new_symbolsdeletion_wordsreplacement_dictrmrnword_graphemes word_diff legal_pronsr\ pron_diffdel_wordr+r+r,replace_symbols$s6         zIpaG2p.replace_symbolscCst|j|dkS)NrW)r`rrvr+r+r,is_unique_in_phoneme_dictWrxz IpaG2p.is_unique_in_phoneme_dictc Cs"t||jd}|j|durt|dfS|jdur)|j|jkr)||dfS|j r8||j vr8||dfS|j dkr-t |dkr| dsN| drd}||j vr||j vr|dd|j vrl|dd}n|dd|j vr|dd}|dur|jr||r|d d vr|j |d d gdfS|d d vr|j |d ddgdfS|j |d dgdfSt |dkr-| d s| dr-d}||j vr||j vr|dd |j vr|dd }n|dd |j vr|dd }|dur-|jr||r-|d d vr!|j |d d gdfS|j |d dgdfS|j dkrgd}gd}t||D]x\}}|d}t |dkr||s]||rd}||j vr||j vrt |}||d|j vr||d}n||d|j vr||d}|dur|jr||rdd|D|j |d dfSq@||j vr|jr||r|j |d dfS|jtkr||j vr||j vr|}|jr||r|j |d dfS|jdur ||dfS||dfS)zXReturns parsed `word` and `status` (bool: False if word wasn't handled, True otherwise).r.NTrz'sz'S)Ttrs)SruɪrOrWrzfr-FR) lcdjmnqurrpuisqulorsqujusqu) rrruʒrrkrrpyiskuloʁskuʒyskrVcSsg|]}|qSr+r+)r)rr+r+r,rZsz)IpaG2p.parse_one_word..F)rr! CHAR_REGEXsearchrBrr6r4rwrrr`endswithrr}rrzip startswithrr) r1rm word_foundcontractions_gcontractions_pcont_gcont_pstarter start_indexr+r+r,parse_one_wordZs     $  * $    zIpaG2p.parse_one_wordtextc s:t|}jdur3zjj|gddd}Wnty2}ztd|dWYd}~nd}~ww|}g}|D]^\}}|rO|fdd|Dq.zK should only have a single item when `without_changes` is False, but found .-) rheteronym_model disambiguate Exceptionrr7r%extendr`rrbrdpop)r1rewords_list_of_tuplernwordswithout_changesrmr\ is_handledsubwords_by_hyphensub_wordrY_r+r0r,__call__s:        zIpaG2p.__call__)T)"__name__ __module__ __qualname__ryr]r^r rr{rr r'r>r?rrrrboolfloatr= staticmethodr8rorArwrrr9rrrr __classcell__r+r+rFr,r"sl     r *@": 83ir)r>r4r] collectionsrtypingrrrrrrr =nemo.collections.common.tokenizers.text_to_speech.ipa_lexiconr Anemo.collections.common.tokenizers.text_to_speech.tokenizer_utilsr r r r$nemo.collections.tts.g2p.models.basernemo.collections.tts.g2p.utilsrrr nemo.utilsrnemo.utils.decoratorsrrr+r+r+r,s $