Module timit

source code

Read tokens, phonemes and audio data from the NLTK TIMIT Corpus.

This corpus contains selected portion of the TIMIT corpus.

• 16 speakers from 8 dialect regions
• 1 male and 1 female from each dialect region
• total 130 sentences (10 sentences per speaker. Note that some sentences are shared among other speakers, especially sa1 and sa2 are spoken by all speakers.)
• total 160 recording of sentences (10 recordings per speaker)
• audio format: NIST Sphere, single channel, 16kHz sampling, 16 bit sample, PCM encoding

Module contents

The timit corpus reader provides 4 functions and 4 data items.

• utterances

  List of utterances in the corpus. There are total 160 utterances, each of which corresponds to a unique utterance of a speaker. Here's an example of an utterance identifier in the list:

     dr1-fvmh0/sx206
       - _----  _---
       | |  |   | |
       | |  |   | |
       | |  |   | `--- sentence number
       | |  |   `----- sentence type (a:all, i:shared, x:exclusive)
       | |  `--------- speaker ID
       | `------------ sex (m:male, f:female)
       `-------------- dialect region (1..8)
  

• speakers

  List of speaker IDs. An example of speaker ID:

     dr1-fvmh0
  

  Note that if you split an item ID with colon and take the first element of the result, you will get a speaker ID.

  >>> itemid = dr1-fvmh0/sx206
  >>> spkrid,sentid = itemid.split('/')
  >>> spkrid
  'dr1-fvmh0'

  The second element of the result is a sentence ID.

• dictionary()

  Phonetic dictionary of words contained in this corpus. This is a Python dictionary from words to phoneme lists.

• spkrinfo()

  Speaker information table. It's a Python dictionary from speaker IDs to records of 10 fields. Speaker IDs the same as the ones in timie.speakers. Each record is a dictionary from field names to values, and the fields are as follows:

   id         speaker ID as defined in the original TIMIT speaker info table
   sex        speaker gender (M:male, F:female)
   dr         speaker dialect region (1:new england, 2:northern,
              3:north midland, 4:south midland, 5:southern, 6:new york city,
              7:western, 8:army brat (moved around))
   use        corpus type (TRN:training, TST:test)
              in this sample corpus only TRN is available
   recdate    recording date
   birthdate  speaker birth date
   ht         speaker height
   race       speaker race (WHT:white, BLK:black, AMR:american indian,
              SPN:spanish-american, ORN:oriental,???:unknown)
   edu        speaker education level (HS:high school, AS:associate degree,
              BS:bachelor's degree (BS or BA), MS:master's degree (MS or MA),
              PHD:doctorate degree (PhD,JD,MD), ??:unknown)
   comments   comments by the recorder
  

The 4 functions are as follows.

• tokenized(sentences=items, offset=False)

  Given a list of items, returns an iterator of a list of word lists, each of which corresponds to an item (sentence). If offset is set to True, each element of the word list is a tuple of word(string), start offset and end offset, where offset is represented as a number of 16kHz samples.

• phonetic(sentences=items, offset=False)

  Given a list of items, returns an iterator of a list of phoneme lists, each of which corresponds to an item (sentence). If offset is set to True, each element of the phoneme list is a tuple of word(string), start offset and end offset, where offset is represented as a number of 16kHz samples.

• audiodata(item, start=0, end=None)

  Given an item, returns a chunk of audio samples formatted into a string. When the fuction is called, if start and end are omitted, the entire samples of the recording will be returned. If only end is omitted, samples from the start offset to the end of the recording will be returned.

• play(data)

  Play the given audio samples. The audio samples can be obtained from the timit.audiodata function.