How to Speak a Language without Knowing It

Can people speak a language they don’t know? Actually, it happens frequently. Travel phrasebooks contain phrases in the speaker’s language (e.g., ‘‘thank you’’) paired with foreign-language translations (e.g., ‘‘ спасибо’’). Since the speaker may not be able to pronounce the foreign-language orthography, phrasebooks additionally provide phonetic spellings that approximate the sounds of the foreign phrase. These spellings employ the familiar writing system and sounds of the speaker’s language. Here is a sample entry from a French phrasebook for English speakers:

The user ignores the French and goes straight to the Franglish. If the Franglish is well designed, an English speaker can pronounce it and be understood by a French listener.

Figure 1 shows a sample entry from another book—an English phrasebook for Chinese speakers. If a Chinese speaker wants to say ‘‘{CJK}UTF8gkai非常感谢你这顿美餐’’, she need only read off the Chinglish “{CJK}UTF8gkai三可 油 否 热斯 弯德否 米欧”, which approximates the sounds of “Thank you for this wonderful meal” using Chinese characters.

Phrasebooks permit a form of accurate, personal, oral communication that speech-to-speech translation devices lack. However, the user is limited to a small set of fixed phrases. In this paper, we lift this restriction by designing and evaluating a software program with the following:

• •

  Input: Text entered by the speaker, in her own language.

• •

  Output: Phonetic rendering of a foreign-language translation of that text, which, when pronounced by the speaker, can be understood by the listener.

The main challenge is that different languages have different orthographies, different phoneme inventories, and different phonotactic constraints, so mismatches are inevitable. Despite this, the system’s output should be both unambiguously pronounceable by the speaker and readily understood by the listener.

Our goal is to build an application that covers many language pairs and directions. The current paper describes a single system that lets a Chinese person speak English.

We take a statistical modeling approach to this problem, as is done in two lines of research that are most related. The first is machine transliteration [3], in which names and technical terms are translated across languages with different sound systems. The other is respelling generation [2], where an English speaker is given a phonetic hint about how to pronounce a rare or foreign word to another English speaker. By contrast, we aim to help people issue full utterances that cross language barriers.

Our system’s input is Chinese. The output is a string of Chinese characters that approximate English sounds, which we call Chinglish. We build several candidate Chinese-to-Chinglish systems and evaluate them as follows:

• •

  We compute the normalized edit distance between the system’s output and a human-generated Chinglish reference.

• •

  A Chinese speaker pronounces the system’s output out loud, and an English listener takes dictation. We measure the normalized edit distance against an English reference.

• •

  We automate the previous evaluation by replace the two humans with: (1) a Chinese speech synthesizer, and (2) a English speech recognizer.

We seek to imitate phonetic transformations found in phrasebooks, so phrasebooks themselves are a good source of training data. We obtained a collection of 1312 Chinese, English, Chinglish$>$ phrasebook tuples ¹¹Dataset can be found at http://www.isi.edu/natural-language/mt/chinglish-data.txt (see Table 1).

We use 1182 utterances for training, 65 for development, and 65 for test. We know of no other computational work on this type of corpus.

Our Chinglish has interesting gross empirical properties. First, because Chinglish and Chinese are written with the same characters, they render the same inventory of 416 distinct syllables. However, the distribution of Chinglish syllables differs a great deal from Chinese (Table 2). Syllables ‘‘si’’ and ‘‘te’’ are very popular, because while consonant clusters like English ‘‘st’’ are impossible to reproduce exactly, the particular vowels in ‘‘si’’ and ‘‘te’’ are fortunately very weak.

We find that multiple occurrences of an English word type are generally associated with the same Chinglish sequence. Also, Chinglish characters do not generally span multiple English words. It is reasonable for ‘‘can I’’ to be rendered as ‘‘kan nai’’, with ‘‘nai’’ spanning both English words, but this is rare.

We model Chinese-to-Chinglish translation with a cascade of weighted finite-state transducers (wFST), shown in Figure 2. We use an online MT system to convert Chinese to an English word sequence (Eword), which is then passed through FST A to generate an English sound sequence (Epron). FST A is constructed from the CMU Pronouncing Dictionary [6].

Next, wFST B translates English sounds into Chinese sounds (Pinyin-split). Pinyin is an official syllable-based romanization of Mandarin Chinese characters, and Pinyin-split is a standard separation of Pinyin syllables into initial and final parts. Our wFST allows one English sound token to map to one or two Pinyin-split tokens, and it also allows two English sounds to map to one Pinyin-split token.

Finally, FST C converts Pinyin-split into Pinyin, and FST D chooses Chinglish characters. We also experiment with an additional wFST E that translates English words directly into Chinglish.

FSTs A, C, and D are unweighted, and remain so throughout this paper.

Our first evaluation (Table 4) is intrinsic, measuring our Chinglish output against references from the test portion of our phrasebook, using edit distance. Here, we start with reference English and measure the accuracy of Pinyin syllable production, since the choice of Chinglish character does not affect the Chinglish pronunciation. We see that the Word-based method has very high accuracy, but low coverage. Our best system uses the Hybrid training/decoding method. As Table 6 shows, the ratio of unseen English word tokens is small, thus large portion of tokens are transformed using word-based method. The average edit distance of phoneme-phrase model and that of hybrid training/decoding model are close, indicating that long phoneme-phrase pairs can emulate word-pinyin mappings.

Our second evaluation is a dictation task. We speak our Chinglish character sequence output aloud and ask an English monolingual person to transcribe it. (Actually, we use a Chinese synthesizer to remove bias.) Then we measure edit distance between the human transcription and the reference English from our phrasebook. Results are shown in Table 7.

Finally, we repeat the last experiment, but removing the human from the loop, using both automatic Chinese speech synthesis and English speech recognition. Results are shown in Table 8. Speech recognition is more fragile than human transcription, so edit distances are greater. Table 5 shows a few examples of the Chinglish generated by the hybrid training and decoding method, as well as the recognized English from the dictation and ASR tasks.

Our work aims to help people speak foreign languages they don’t know, by providing native phonetic spellings that approximate the sounds of foreign phrases. We use a cascade of finite-state transducers to accomplish the task. We improve the model by adding phrases, word boundary constraints, and improved alignment.

In the future, we plan to cover more language pairs and directions. Each target language raises interesting new challenges that come from its natural constraints on allowed phonemes, syllables, words, and orthography.