Contextual multi-channel speech to text

A system for improving a transcription, comprising:

• a computer processor;
• a repository storing the transcription, wherein the transcription comprises a first plurality of tokens of a first utterance of a reliable channel and a first token of an utterance of an unreliable channel;
• and
• a transcription improver executing on the computer processor and configured to:
• identify, in the transcription, the first plurality of tokens and the first token,
• generate, using a machine learning model, a first vector embedding for the first token and a first plurality of vector embeddings for the first plurality of reliable channel tokens,
• calculate a plurality of vector distances between the first vector embedding and the first plurality of vector embeddings,
• generate, for the first token and using the plurality of vector distances, a score corresponding to a second token of the first plurality of tokens,
• determine the score is within a threshold score, and
• in response to determining that the score is within the threshold score, replace, in the transcription, the first token with the second token.

A method for improving a transcription, comprising:

• identifying, in the transcription, a first plurality of tokens of a preceding utterance of a reliable channel, a second plurality of tokens of an utterance of an unreliable channel, and a third plurality of tokens of a proceeding utterance of the reliable channel;
• generating, for a first token of the second plurality of tokens and using a machine learning model, a plurality of alternative utterances each comprising a plurality of alternative tokens, wherein the first token corresponds to a score;
• generating, using the machine learning model and for each respective alternative utterance in the plurality of alternative utterances:
• a consequent probability that the respective alternative utterance follows the preceding utterance of the reliable channel, and
• a precedent probability that the respective alternative utterance precedes the proceeding utterance of the reliable channel;
• and
• substituting, based at least in part on the consequent probability, the precedent probability, and the score, an alternative utterance from the plurality of alternative utterances as a replacement for the utterance of the unreliable channel.