Speech model migration in Q1 2024

Migration status

As of July 2024, no projects have been automatically migrated. The automatic migration will remain on hold until further notice (no ETA).

You can continue to use the classic Speech-to-Text models but are encouraged to opt in to use conformer-based speech models.

Migration overview

For certain language tag and speech setting combinations, the Conversational Agents (Dialogflow CX) and Dialogflow ES traffic with audio will gradually route away from the classic Speech-to-Text models behind the command_and_search, default, phone_call, and video model identifiers to the new conformer-based speech models. As the redirection will be internal only, the current model identifiers and code will remain valid and continue to serve traffic.

The tables in the Migration for en and en-uslanguage tags and Migration for de, en-au, en-gb, en-in, es, es-es, es-us, fr, fr-ca, fr-fr, it, ja, nl, pt-br language tags sections show the routing of Speech-to-Text models that will take place when the migration takes effect. If your configuration matches the language tag and the first two columns in the table, the new model identifier is listed in the third column. No change is expected for other configurations.

See the Migration options section to choose the best option for you.

Migration for en and en-us language tags

Current model identifier	Current single_utterance setting *	Model identifier that traffic will be directed to
command_and_search	false	latest_long
command_and_search	true	latest_short
default	false	telephony
phone_call	false	telephony
phone_call (model_variant is not USE_STANDARD)	true	telephony_short
video	false	telephony
Unspecified	true/false	Model is auto-selected from among the above, based on the parameters in InputAudioConfig

Symbol	Description
*	If your Conversational Agents (Dialogflow CX) agent has enabled advanced timeout-based end of speech sensitivity, then the model identifier is routed assuming single_utterance is false, even if the setting is set to true at runtime.

Migration for de, en-au, en-gb, en-in, es, es-es, es-us, fr, fr-ca, fr-fr, it, ja, nl, pt-br language tags

Current model identifier	Current single_utterance setting *	Model identifier that traffic will be directed to
command_and_search	false	latest_long
command_and_search	true	latest_short
default	false	latest_long
Unspecified	true/false	Model is auto-selected from among the above, based on the parameters in InputAudioConfig

Symbol	Description
*	If your Conversational Agents (Dialogflow CX) agent has enabled advanced timeout-based end of speech sensitivity, then the model identifier is routed assuming single_utterance is false, even if the setting is set to true at runtime.

Migration options

There are three migration options: early opt-in, automatic, and temporary opt-out.

Opt in and migrate earlier

To opt in proactively, change the speech model identifier that you have been using in the Conversational Agents (Dialogflow CX) API and Dialogflow ES API with the updated one. For Conversational Agents (Dialogflow CX), you can also change the model identifier in the Advanced speech settings under Speech and IVR settings in the Conversational Agents (Dialogflow CX) console.

Speech model specified in the API at runtime takes precedence over the speech model specified in the Conversational Agents (Dialogflow CX) console, unless you enable the Override request-level speech model setting. Read more about precedence in speech models.

For partner telephony integrations, check with your telephony provider if the speech model is being set in the API. If you are using the Conversational Agents (Dialogflow CX) Phone Gateway, review the speech model limitations.

Migrating your project proactively gives you time to test the models and take advantage of the improved accuracy and robustness earlier. This approach is highly recommended, especially if you are using a model that will migrate to the new telephony and telephony_short models. Most integrations require single_utterance = true, so if you are not using advanced timeout-based end of speech sensitivity you'll likely want to use telephony_short or latest_short.

If you experience issues with the new models, open a Google Cloud support case with issue details.

Migrate automatically

To be migrated automatically, no action is required on your part. The existing Speech-to-Text model identifiers are valid and we will shift traffic internally.

Opt out and migrate later

Due to the suspended migration status, new opt-out requests are not accepted. If your projects were opted out before July 2024, they will remain in this state until further announcement.