The "latest" model tags in the Speech-to-Text API give access to two new model tags that can be used when you specify the model field. These models are designed to give you access to the latest speech technology and machine learning research from Google, and can provide higher accuracy for speech recognition over other available models. However, some features that are supported by other available models are not yet supported by the "latest" models.
The latest models are based on the Conformer Speech Model technology from Google. To find out more, see Google Research Publications.
Using the latest models requires a general understanding of using the Speech-to-Text API or UI. Please see our Quickstarts if this is your first time using it.
The latest models are available in two different versions:
latest_shortmodel is for short utterances that are a few seconds in length. It is useful for trying to capture commands or other single shot directed speech use cases. When you use this model, the service will stop transcribing audio after the first utterance is detected and completed.
Separate recognition per channel is not supported by this model. While multi channel audio is accepted, only the first channel will be processed and transcribed.
latest_longmodel is for any kind of long form content such as media or spontaneous speech and conversations.
The goal of the latest models is to bring the latest in speech technology directly to Google Cloud users. Our current Latest models are based on the Conformer Speech Model technology from Google, but this may change in the future. To find out more, check out Google Research Publications list.
latest_short models are billed as "Standard".
For more information, see Pricing.
Latest models are based on rapidly advancing machine learning technology. For this reason we might perform model updates or refreshes more frequently than on our other models. These updates can add additional features or make slight changes to accuracy or latency.
Latest models are available in more than 20 languages and more than 50 variants. We are always adding languages, so refer to Languages for the most up to date list.
Feature support and limitations
Feature support varies by language. See Languages for a full list of supported features.
The latest models do not currently support the following features:
Confidence Scores - The API will return a value, but it is not truly a confidence score.
Biasing - Only the
latest_shortmodel supports biasing.
Diarization - Neither latest model supports diarization.