Managing access to sensitive data in transcriptions

A system comprising:

• a first one or more electronic devices to implement a machine learning service in a provider network;
• and
• a second one or more electronic devices to implement a transcription service in the provider network, the transcription service including instructions that upon execution cause the transcription service to:
• receive, via a first application programming interface offered by the transcription service, a first request to generate a redacted transcript of content;
• obtain a transcript of the content;
• send the transcript to a model endpoint hosted by the machine learning service, the model endpoint capable of identifying sensitive entities in the transcript;
• receive an inference response identifying one or more sensitive entities in the transcript;
• generate the redacted transcript based at least on the transcript and the inference response by replacing each sensitive entity of the one or more sensitive entities with a respective sensitive entity tag, the respective sensitive entity tag comprising a name indicating a sensitive entity data type of the replaced sensitive entity;
• receive, via a second application programming interface offered by the transcription service, a second request for all sensitive entities of a particular sensitive entity data type, the second request comprising a name of the particular sensitive entity data type;
• query a representation of the one or more sensitive entities based on the name of the particular sensitive entity data type;
• and
• return, via the second application programming interface, all sensitive entities, of the one or more sensitive entities, that are of the particular sensitive entity data type.

A computer-implemented method comprising:

• receiving, by a transcription service in a provider network, via a first application programming interface offered by the transcription service, a first request to generate a redacted transcript of audio content, wherein the transcription service is implemented by one or more electronic devices;
• generating, by the transcription service, a transcript of the audio content;
• postprocessing, by the transcription service, the transcript to generate a postprocessed transcript, wherein the postprocessing includes number normalization and punctuation;
• sending, by the transcription service, the postprocessed transcript to a model endpoint hosted by a machine learning service in the provider network, wherein the model endpoint hosts a plurality of machine learning models, wherein each machine learning model of the plurality of machine learning models is trained to identify sensitive entities of a different respective sensitive entity data type, wherein the postprocessed transcript is received at the model endpoint, wherein the machine learning service is implemented by one or more electronic devices;
• receiving, by the transcription service, an inference response identifying a plurality of sensitive entities in the postprocessed transcript, wherein the inference response identifies each sensitive entity of the plurality of sensitive entities by a respective sensitive entity data type and a respective byte offset;
• generating, by the transcription service, the redacted transcript using the postprocessed transcript and the inference response, wherein the redacted transcript comprises a respective sensitive entity tag in place of each sensitive entity of the plurality of sensitive entities, wherein the respective sensitive entity tag comprises a name indicating the respective sensitive entity data type of the sensitive entity;
• receiving, by the transcription service, via a second application programming interface offered by the transcription service, a second request to obtain all sensitive entities of a particular sensitive entity data type, wherein the second request comprises a name of the particular sensitive entity data type;
• querying, by the transcription service, a representation of the plurality of sensitive entities based on the name of the particular sensitive entity data type;
• and
• returning, by the transcription service, via the second application programming interface, all sensitive entities, of the plurality of sensitive entities, that are of the particular sensitive entity data type.