Creating a custom acoustic model
To customize the acoustic model to a particular domain, a collection of speech data is required. This collection consists of a set of audio files of speech data, and a text file of transcriptions of each audio file. The audio data should be representative of the scenario in which you would like to use the recognizer.
- If you would like to better recognize speech in a noisy factory environment, the audio files should consist of people speaking in a noisy factory.
- If you are interested in optimizing performance for a single speaker, e.g. you would like to transcribe all of FDR’s Fireside Chats, then the audio files should consist of many examples of that speaker only.
Preparing your data to customize the acoustic model
An acoustic data set for customizing the acoustic model consists of two parts: (1) a set of audio files containing the speech data and (2) a file containing the transcriptions of all audio files.
Audio Data Recommendations
- All audio files in the data set should be stored in the WAV (RIFF) audio format.
- The audio must have a sampling rate of 8 kHz or 16 kHz and the sample values should be stored as uncompressed PCM 16-bit signed integers (shorts).
- Only single channel (mono) audio files are supported.
- The audio files must be between 100ms and 1 minute in length. Each audio file should ideally start and end with at least 100ms of silence, and somewhere between 500ms and 1 second is common.
- If you have background noise in your data, it is recommended to also have some examples with longer segments of silence, e.g. a few seconds, in your data, before and/or after the speech content.
- Each audio file should consist of a single utterance, e.g. a single sentence for dictation, a single query, or a single turn of a dialog system.
- Each audio file to in the data set should have a unique filename and the extension “wav”.
- The set of audio files should be placed in a single folder without subdirectories and the entire set of audio files should be packaged as a single ZIP file archive.
Data imports via the web portal are currently limited to 2 GB, so this is the maximum size of an acoustic data set. This corresponds to approximately 17 hours of audio recorded at 16 kHz or 34 hours of audio recorded at 8 kHz. The main requirements for the audio data are summarized in the following table.
|File Format||RIFF (WAV)|
|Sampling Rate||8000 Hz or 16000 Hz|
|Sample Format||PCM, 16 bit integers|
|File Duration||0.1 seconds < duration < 60 seconds|
|Silence Collar||> 0.1 seconds|
|Maximum Archive Size||2 GB|
The transcriptions for all WAV files should be contained in a single plain-text file. Each line of the transcription file should have the name of one of the audio files, followed by the corresponding transcription. The file name and transcription should be separated by a tab (\t).
speech01.wav speech recognition is awesome speech02.wav the quick brown fox jumped all over the place speech03.wav the lazy dog was not amused
The transcriptions will be text-normalized so they can be processed by the system. However, there are some very important normalizations that must be done by the user prior to uploading the data to the Custom Speech Service. Please consult the section on transcription guidelines for the appropriate language when preparing your transcriptions.
The following steps are done using the Custom Speech Service Portal.
Step 1: Importing the acoustic data set
Once the audio files and transcriptions have been prepared, they are ready to be imported to the service web portal.
To do so, first ensure you are signed into the system. Then click the “Custom Speech” drop-down menu on the top ribbon and select “Adaptation Data”. If this is your first time uploading data to the Custom Speech Service, you will see an empty table called “Acoustic Datasets”. The current locale is reflected in the table title. If you would like to import acoustic data of a different language, click on “Change Locale”. Additional information on supported languages can be found in the section on changing locale.
Click the “Import New” button, located directly below the table title and you will be taken to the page for uploading a new data set.
Enter a Name and Description in the appropriate text boxes. These are useful for keeping track of various data sets you upload. Next, click “Choose File" for the “Transcription File” and “WAV files” and select your plaint-text transcription file and zip archive of WAV files, respectively. When this is complete, click “Import” to upload your data. Your data will then be uploaded. For larger data sets, this may take several minutes.
When the upload is complete, you will return to the "Acoustic Datasets" table and will see an entry that corresponds to your acoustic data set. Notice that it has been assigned a unique id (GUID). The data will also have a status that reflects its current state. Its status will be “Waiting” while it is being queued for processing, “Processing” while it is going through validation, and “Complete” when the data is ready for use.
Data validation includes a series of checks on the audio files to verify the file format, length, and sampling rate, and on the transcription files to verify the file format and perform some text normalization.
When the status is “Complete” you can click “Details” to see the acoustic data verification report. The number of utterances that passed and failed verification will be shown, along with details about the failed utterances. In the example below, two WAV files failed verification because of improper audio format (in this data set, one had an incorrect sampling rate and one was the incorrect file format).
At some point, if you would like to change the Name or Description of the data set, you can click the “Edit” link and change these entries. Note that you cannot modify the audio files or transcriptions.
Step 2: Creating a custom acoustic model
Once the status of your acoustic data set is “Complete”, it can be used to create a custom acoustic model. To do so, click “Acoustic Models” in the “Custom Speech” drop-down menu. You will see a table called "Your models” that lists all of your custom acoustic models. This table will be empty if this is your first use. The current locale is shown in the table title. Currently, acoustic models can be created for US English only.
To create a new model, click “Create New” under the table title. As before, enter a name and description to help you identify this model. For example, the "Description"" field can be used to record which starting model and acoustic data set were used to create the model. Next, select a “Base Acoustic Model” from the drop-down menu. The base model is the model which is the starting point for your customization. There are two base acoustic models to choose from. The Microsoft Search and Dictation AM is appropriate for speech directed at an application, such as commands, search queries, or dictation. The Microsoft Conversational model is appropriate for recognizing speech spoken in a conversational style. This type of speech is typically directed at another person and occurs in call center or meetings. Note that latency for partial results in Conversational models is higher than in Search and Dictation models.
Next, select the acoustic data you wish to use to perform the customization using the drop-down menu.
You can optionally choose to perform offline testing of your new model when the processing is complete. This will run a speech-to-text evaluation on a specified acoustic data set using the customized acoustic model and report the results. To perform this testing, select the “Accuracy Testing” check box. Then select a language model from the drop-down menu. If you have not created any custom language models, only the base language models will be in the drop-down list. Please see the description of the base language models in the guide and select the one that is most appropriate.
Finally, select the acoustic data set you would like to use to evaluate the custom model. If you perform accuracy testing, it is important to select an acoustic data that is different from the one used for the model creation to get a realistic sense of the model’s performance. Also note that offline testing is limited to 1000 utterances. If the acoustic dataset for testing is larger than that, only the first 1000 utterances will be evaluated.
When you are ready to start running the customization process, press “Create”.
You will now see a new entry in the acoustic models table corresponding to this new model. The status of the process is reflected in the table. The status states are “Waiting”, “Processing” and “Complete”.