Utterly Voice depends on third-party software for speech recognition. You can configure up to two recognizers: one for short utterances and one for long utterances, switched automatically based on a speech duration threshold. You can choose the same or different recognizers for short and long utterances from the following options:
General application settings are found in the
config\settings.yaml
file found in the application directory.
This file has the following recognizer settings:
| Setting | Type | Description |
|---|---|---|
recognizerShort
|
String |
The recognizer to use for short utterances
(active speech duration shorter than
recognizerThresholdMilliseconds).
This value can be "vosk", "whispercpp", "azure", "deepgram",
"google_v1", or "google_v2".
|
recognizerLong
|
String |
The recognizer to use for long utterances
(active speech duration at or above
recognizerThresholdMilliseconds).
This value can be "vosk", "whispercpp", "azure", "deepgram",
"google_v1", or "google_v2".
You can choose the same or different recognizers for short and long
utterances.
|
recognizerThresholdMilliseconds
|
Number |
Active speech duration threshold in milliseconds.
Utterances shorter than this threshold use
recognizerShort;
utterances at or above this threshold use
recognizerLong.
|
voskRecognizerConfig
|
Map Collection | Settings specific for the Vosk recognizer. See Vosk below. |
whisperCppRecognizerConfig
|
Map Collection | Settings specific for the Whisper recognizer. See Whisper below. |
azureRecognizerConfig
|
Map Collection | Settings specific for the Microsoft Azure recognizer. See Microsoft Azure below. |
deepgramRecognizerConfig
|
Map Collection | Settings specific for the Deepgram recognizer. See Deepgram below. |
googleV1RecognizerConfig
|
Map Collection | Settings specific for the Google Speech-to-Text V1 recognizer. See Google Cloud Speech-to-Text below. |
googleV2RecognizerConfig
|
Map Collection | Settings specific for the Google Speech-to-Text V2 recognizer. See Google Cloud Speech-to-Text below. |
Vosk is a free offline speech recognizer. The default settings use this recognizer for short utterances. It runs on your local machine, so it does not require an internet connection. This recognizer has one of the lowest latencies of any recognizer, and processing is performed on your CPU.
If you use this speech recognizer, Utterly Voice will use about 5 GB of memory for the recognizer. Your machine should have at least 8 GB of memory, ideally 16 GB or more.
Here are the configuration settings:
| Setting | Type | Description |
|---|---|---|
model
|
String |
The path to the model directory
relative to the application directory or an absolute path.
Utterly Voice provides a recommended model,
but you can try other models.
Download a model
that you want to try,
unzip the file,
and update this setting to be the path
to the model directory that contains a
README file in it.
|
Whisper is a free offline speech recognizer. The default settings use this recognizer for long utterances. It runs on your local machine, so it does not require an internet connection.
Whisper does not provide any effective means to control speech recognition bias. This means that the bias settings you apply in your settings files do not have any effect with this recognizer.
Whisper is not designed for realtime streaming audio processing. It is primarily designed for processing audio files. Due to this, the recognition latency is larger than it is for other recognizers. Whisper uses your GPU for processing, so you may notice more latency if you do not have a powerful GPU.
If you use this speech recognizer, Utterly Voice will use less than 1 GB of memory for the recognizer.
Here are the configuration settings:
| Setting | Type | Description |
|---|---|---|
model
|
String |
The path to the model file
relative to the application directory or an absolute path.
This defaults to '..\models\whispercpp\ggml-base.en.bin'.
If you have a high-end GPU,
you can change this to
'..\models\whispercpp\ggml-large-v3-turbo.bin',
which already exists in that folder.
You can also download and try other models from
Hugging Face
(*.bin files)
and update this setting to be the path to the model file.
|
Microsoft Azure Speech to Text is a paid online speech recognition service. This recognizer runs on Microsoft servers, so you need to have an active internet connection to use this option.
If you have an accent, this is currently the best recognizer we provide. In the language settings below, you can set the English locale.
You can optionally create a custom speech model, using your voice and jargon. Instructions are below.
On rare occasions, requests sent to this service receive error responses. If this happens, turn the microphone off for one minute, which gives the service time to recover.
To set this up, you need to create an account, create a speech resource, select a region, and copy a secret key for authentication:
eastus is the region identifier for the
https://eastus.api.cognitive.microsoft.com/ endpoint.
You will need this region identifier below.
If you would like to create a custom speech model, follow the instructions below. Note that hosting a custom model on Azure will result in monthly charges, even if you are not actively using the model. See the Azure pricing for details.
audioFiles setting in the general settings file
to be the number of utterances you want to capture.
The text transcription file should contain one line for each utterance.
Each line contains the audio file name, followed by a tab character,
followed by the utterance text.
For example: utterance-25.wav	eurasian wigeon.customModelIdentifier
setting described below.
Here are the Utterly Voice configuration settings:
| Setting | Type | Description |
|---|---|---|
secretKey
|
String | The resource key you acquired above. |
region
|
String | The region identifier you acquired above. |
language
|
String | The language used for recognition. This must be an English-based language. There are many locale-specific English options for improved recognition with accents ("en-AU", "en-GB", "en-HK", etc.). See the Azure Speech language options. |
customModelIdentifier
|
String | The custom speech model identifier. Azure calls this the Endpoint ID. If you are not using a custom speech model, this should be set to an empty string. |
Deepgram is a paid online speech recognition service. This recognizer runs on Deepgram servers, so you need to have an active internet connection to use this option. To set this up:
Here are the Utterly Voice configuration settings:
| Setting | Type | Description |
|---|---|---|
secretKey
|
String | The API key you acquired above. |
model
|
String | The recognizer model. The recommended model is provided in the default settings, but you are free to experiment with other models. You must choose a model that supports streaming (Whisper does not support streaming). |
alternatives
|
Number | The number of transcript alternatives sent by the service. These alternatives are used to apply bias to select the best transcript. Some of the Deepgram models only work if this is set to 1. This is not documented by Deepgram. The "enhanced-general" model can be set for 3 alternatives. The "nova-2" model must be set for 1 alternatives. For all other models, you should experiment with either 1 (Utterly Voice bias settings have no effect) or 3 (Utterly Voice bias settings work as expected). |
minConfidence
|
Number |
This recognizer frequently interprets noise as text.
Fortunately, the service usually returns a low confidence score in these cases.
This setting allows you to set a minimum confidence score.
If the score is below this value,
the recognized text will be ignored.
You can see the confidence scores for recognized text in the
log.txt file in the application directory.
|
Google Cloud Speech-to-Text is a paid online speech recognition service. This recognizer runs on Google servers, so you need to have an active internet connection to use this option. To set this up, see Google's setup instructions and complete the following steps:
secret_google_credentials.json.
Be sure to keep this file in a secure location.
If anyone copies this file,
they can make calls to Google Cloud services
that you will be billed for.
If you work in an organization that prevents downloading of service account keys, you can use the gcloud command line tool instead. Follow the steps to install this tool. Once installed, you can use the gcloud auth application-default login command with no command line flags to authenticate calls from Utterly Voice. This command will have you follow some steps to create a credentials file on your computer in a location that Utterly Voice can find. Unless you delete this file, you only need to call this command once.
The service is available in either version V1
(google_v1) or V2 (google_v2).
Here are the Utterly Voice configuration settings:
| Setting | Type | Description |
|---|---|---|
model
|
String | The recognizer model (V1 and V2). The recommended model is provided in the default settings, but you are free to experiment with the latest fast models or slower generative models. |
project
|
String | The Google Cloud Project ID you created in steps above (V2 only). |
location
|
String | The chosen location of the service (V2 only). The default location is central US, but you can change this to another location that may be closer to your geographic location. Note that Google restricts certain models to specific locations, and these are not well documented, so you may need to experiment with model/location combinations. |
recognizer
|
String | This field is only used when you want to create a custom model (V2 only). The default value of "_" specifies to use the default model. We have not tested custom models, but you can contact us if you want to try this feature and are having issues. |