The CMU Pronouncing Dictionary (also known as cmudict) is a machine-readable pronunciation dictionary for North American English. It was created at Carnegie Mellon University and is maintained by the university's Speech Group. The dictionary provides a mapping from words to their phonetic transcriptions, using a set of 39 phonemes (plus stress markers). It is released into the public domain, making it a foundational resource for speech technology and computational linguistics.
The dictionary's origins trace back to the 1980s, developed for use in speech recognition research at Carnegie Mellon. Its first public release occurred in 1993, and it has been continuously updated since. The current version (0.7b) contains over 134,000 entries, covering common words, proper nouns, and acronyms. Each entry lists the word in lowercase, followed by its phoneme sequence with stress indicators (0, 1, or 2) placed after vowels.
Phoneme Set and Transcription Format
The CMU Pronouncing Dictionary uses a phoneme set based on the ARPABET system, originally developed for the ARPA Speech Understanding Research project. This set includes 39 phonemes, such as vowels (e.g., AA, IY, UW) and consonants (e.g., K, S, T). Stress is marked with digits: 0 for no stress, 1 for primary stress, and 2 for secondary stress. For example, the word "hello" is transcribed as "HH AH0 L OW1". This format is simple, ASCII-based, and easily parsed by software.
The dictionary also includes multiple pronunciations for many words, indicated by parenthetical numbers (e.g., "the(1)" and "the(2)"). These variants capture regional or contextual differences, such as the schwa versus stressed vowel in "the". Additionally, it contains entries for inflected forms, compounds, and some foreign words commonly used in English.
Applications in Speech Technology
The dictionary is a standard resource in Artificial intelligence and Machine learning pipelines for speech processing. It is used as a pronunciation lexicon in automatic speech recognition (ASR) systems, enabling models to map acoustic signals to word sequences. In text-to-speech (TTS) synthesis, it provides the phonetic input needed to generate natural-sounding speech. Many open-source toolkits, such as Kaldi and ESPnet, include cmudict as a default lexicon.
Beyond ASR and TTS, the dictionary is used in Natural language processing tasks like grapheme-to-phoneme conversion, where models learn to predict pronunciations for unseen words. It also serves as a benchmark for evaluating pronunciation prediction algorithms. Researchers at institutions like Stanford AI Lab and MIT CSAIL have used cmudict in studies on phonetic alignment and speech synthesis.
Format and Distribution
The dictionary is distributed as a plain text file, with one entry per line. The format is: word followed by one or more spaces, then the phoneme sequence. Comments and metadata are minimal, keeping the file compact. It is available for download from the Carnegie Mellon Speech Group's website and is mirrored on many open-source repositories. The public-domain status allows unrestricted use, modification, and redistribution, making it a preferred choice for commercial and academic projects alike.
Over the years, several derived resources have been created from cmudict, including pronunciation dictionaries for other languages (via translation) and lexicons with added syllable boundaries or part-of-speech tags. These derivatives are often used in Deep learning models for speech and language, where they provide supervision signals for end-to-end systems.
Limitations and Extensions
While comprehensive, the dictionary has known limitations. It covers only North American English, and its phoneme set does not capture all dialectal variations. Some entries may be outdated or missing recent proper nouns and technical terms. To address this, community efforts have produced extended versions, such as cmudict with added words from Wiktionary or the CMUdict with stress and syllable annotations. These extensions are frequently used in modern Transformer (architecture)-based TTS systems, which require large, diverse lexicons.
Despite these limitations, the CMU Pronouncing Dictionary remains a cornerstone of speech research. Its simplicity, reliability, and open license have ensured its continued relevance for over three decades, from early rule-based systems to contemporary Neural network approaches.
See Also
- Sequence-to-Sequence (Seq2Seq) for models that use pronunciation lexicons
- Data Augmentation for techniques that generate synthetic speech data
- Carnegie Mellon University as the origin institution