How to create human-labeled transcriptions

Human-labeled transcriptions are word-by-word transcriptions of an audio file. You use human-labeled transcriptions to improve recognition accuracy, especially when words are deleted or incorrectly replaced.

A large sample of transcription data is required to improve recognition. We suggest providing between 1 and 20 hours of transcription data. The Speech service will use up to 20 hours of audio for training. On this page, we'll review guidelines designed to help you create high-quality transcriptions. This guide is broken up by locale, with sections for US English, Mandarin Chinese, and German.

US English (en-US)

Human-labeled transcriptions for English audio must be provided as plain text, only using ASCII characters. Avoid the use of Latin-1 or Unicode punctuation characters. These characters are often inadvertently added when copying text from a word-processing application or scraping data from web pages. If these characters are present, make sure to update them with the appropriate ASCII substitution.

Here are a few examples:

Text normalization for US English

Text normalization is the transformation of words into a consistent format used when training a model. Some normalization rules are applied to text automatically, however, we recommend using these guidelines as you prepare your human-labeled transcription data:

• Write out abbreviations in words.
• Write out non-standard numeric strings in words (such as accounting terms).
• Non-alphabetic characters or mixed alphanumeric characters should be transcribed as pronounced.
• Abbreviations that are pronounced as words shouldn't be edited (such as "radar", "laser", "RAM", or "NATO").
• Write out abbreviations that are pronounced as separate letters with each letter separated by a space.
• If you use audio, transcribe numbers as words that match the audio (for example, "101" could be pronounced as "one oh one" or "one hundred and one").
• Avoid repeating characters, words, or groups of words more than three times, such as "yeah yeah yeah yeah". Lines with such repetitions might be dropped by the Speech service.

Here are a few examples of normalization that you should perform on the transcription:

The following normalization rules are automatically applied to transcriptions:

• Use lowercase letters.
• Remove all punctuation except apostrophes within words.
• Expand numbers into words/spoken form, such as dollar amounts.

Here are a few examples of normalization automatically performed on the transcription:

Mandarin Chinese (zh-CN)

Human-labeled transcriptions for Mandarin Chinese audio must be UTF-8 encoded with a byte-order marker. Avoid the use of half-width punctuation characters. These characters can be included inadvertently when you prepare the data in a word-processing program or scrape data from web pages. If these characters are present, make sure to update them with the appropriate full-width substitution.

Here are a few examples:

Text normalization for Mandarin Chinese

Text normalization is the transformation of words into a consistent format used when training a model. Some normalization rules are applied to text automatically, however, we recommend using these guidelines as you prepare your human-labeled transcription data:

• Write out abbreviations in words.
• Write out numeric strings in spoken form.

Here are a few examples of normalization that you should perform on the transcription:

The following normalization rules are automatically applied to transcriptions:

• Remove all punctuation
• Expand numbers to spoken form
• Convert full-width letters to half-width letters
• Using uppercase letters for all English words

Here are some examples of automatic transcription normalization:

German (de-DE) and other languages

Human-labeled transcriptions for German audio (and other non-English or Mandarin Chinese languages) must be UTF-8 encoded with a byte-order marker. One human-labeled transcript should be provided for each audio file.

Text normalization for German

Text normalization is the transformation of words into a consistent format used when training a model. Some normalization rules are applied to text automatically, however, we recommend using these guidelines as you prepare your human-labeled transcription data:

• Write decimal points as "," and not ".".
• Write time separators as ":" and not "." (for example: 12:00 Uhr).
• Abbreviations such as "ca." aren't replaced. We recommend that you use the full spoken form.
• The four main mathematical operators (+, -, *, and /) are removed. We recommend replacing them with the written form: "plus," "minus," "mal," and "geteilt."
• Comparison operators are removed (=, <, and >). We recommend replacing them with "gleich," "kleiner als," and "grösser als."
• Write fractions, such as 3/4, in written form (for example: "drei viertel" instead of 3/4).
• Replace the "€" symbol with its written form "Euro."

Here are a few examples of normalization that you should perform on the transcription:

The following normalization rules are automatically applied to transcriptions:

• Use lowercase letters for all text.
• Remove all punctuation, including various types of quotation marks ("test", 'test', "test„, and «test» are OK).
• Discard rows with any special characters from this set: ¢ ¤ ¥ ¦ § © ª ¬ ® ° ± ² µ × ÿ ج¬.
• Expand numbers to spoken form, including dollar or Euro amounts.
• Accept umlauts only for a, o, and u. Others will be replaced by "th" or be discarded.

Here are a few examples of normalization automatically performed on the transcription:

Text normalization for Japanese

In Japanese (ja-JP), there's a maximum length of 90 characters for each sentence. Lines with longer sentences will be discarded. To add longer text, insert a period in between.

Next Steps

• Prepare and test your data
• Inspect your data
• Evaluate your data
• Train your model
• Deploy your model