Machine learning-based speech-to-text transcription cloud intermediary

1. A computer-implemented method of cloud-based speech recognition from audio channels without prior training to adapt to speaker(s) in the audio channels, the method including:

accessing a machine-learning engine;

submitting hundreds of speech samples to multiple speech recognition engines, and analyzing how transcription error rates of the speech recognition engines vary with sound and speech characteristics of the speech samples;

receiving an audio channel and qualifying the speech recognition engines as capable of transcribing the audio channel and/or its parts, taking into account at least a recording codec of the audio channel, available transcoding from the recording codec to a speech recognition engine supported codec, a length of the audio channel, and a language of the audio channel;

applying an audio channel analyzer to the audio channel to characterize audio fidelity, background noise, concurrent speech by multiple speakers, timbre, pitch, and audio distortion of the audio channel; and

using the machine learning engine, selecting a speech recognition engine that is qualified as capable of transcribing the audio channel and/or its parts or a transcoded version of the audio channel and/or its parts, based on a combination of least three of the following characteristics:

audio fidelity;

background noise;

concurrent speech by multiple users;

timbre;

pitch; and

audio distortion of the audio channel.

2. The computer-implemented method of claim 1 , further including:

using the multiple speech recognition engines on the audio channel, including using the speech recognition engines sequentially when a first speech recognition engine reports a low confidence score on some or all of its transcription.

3. The computer-implemented method of claim 1 , further including:

using the multiple speech recognition engines on all or separate parts of the audio channel, including using the speech recognition engines when voting on transcription results is used, when different speakers on different tracks of the audio channel, and when different speakers take turns during segments of the audio channel.

4. The computer-implemented method of claim 1 , further including applying the method to separation/identification (diarization) engines.

5. The computer-implemented method of claim 1 , further including applying the method to auto-punctuation engines.

6. The computer-implemented method of claim 1 , further including applying a silence analyzer to the speech samples and the audio channel prior to submission to parse out silent parts of speech.

7. The computer-implemented method of claim 1 , further including performing testing periodically, including daily, weekly, or monthly, wherein the testing comprises the submitting of hundreds of speech samples to multiple speech recognition engines, and the analyzing of how transcription error rates of the speech recognition engines vary with sound and speech characteristics of the speech samples.

8. A computer-implemented method of cloud-based speech recognition from audio channels without prior training to adapt to speaker(s) in the audio channels, the method including:

accessing a machine-learning engine;

submitting thousands of speech samples to multiple speech recognition engines, and analyzing how transcription error rates of the speech recognition engines vary with sound and speech characteristics of the speech samples;

receiving an audio channel and qualifying the speech recognition engines as capable of transcribing the audio channel and/or its parts, taking into account at least a recording codec of the audio channel, available transcoding from the recording codec to a speech recognition engine supported codec, a length of the audio channel, and a language of the audio channel;

applying an audio channel analyzer to the audio channel to characterize audio fidelity, background noise, concurrent speech by multiple speakers, timbre, pitch, and audio distortion of the audio channel; and

using the machine learning engine, selecting a speech recognition engine that is qualified as capable of transcribing the audio channel and/or its parts or a transcoded version of the audio channel and/or its parts, based on a combination of at least three of the following characteristics:

audio fidelity;

background noise;

concurrent speech by multiple users;

timbre;

pitch; and

audio distortion of the audio channel.

9. A computer-implemented method of cloud-based speech recognition from audio channels without prior training to adapt to speaker(s) in the audio channels, the method including:

accessing a machine-learning engine;

submitting dozens of speech samples to multiple speech recognition engines, and analyzing how transcription error rates of the speech recognition engines vary with sound and speech characteristics of the speech samples;

receiving an audio channel and qualifying the speech recognition engines as capable of transcribing the audio channel and/or its parts, taking into account at least a recording codec of the audio channel, available transcoding from the recording codec to a speech recognition engine supported codec, a length of the audio channel, and a language of the audio channel;

applying an audio channel analyzer to the audio channel to characterize audio fidelity, background noise, concurrent speech by multiple speakers, timbre, pitch, and audio distortion of the audio channel; and

using the machine learning engine, selecting a speech recognition engine that is qualified as capable of transcribing the audio channel and/or its parts or a transcoded version of the audio channel and/or its parts, based on a combination of at least three of the following characteristics:

audio fidelity;

background noise;

concurrent speech by multiple users;

timbre;

pitch; and

audio distortion of the audio channel.

10. A system for cloud-based speech recognition from audio channels without prior training to adapt to speaker(s) in the audio channels, comprising:

a machine-learning engine trained to process speech samples;

an analyzer for submitting and analyzing submitting hundreds of speech samples to multiple speech recognition engines, the analyzer determining error rates of the speech recognition engines based on sound and speech characteristics of the speech samples;

the analyzer including receiving an audio channel and qualifying the speech recognition engines as capable of transcribing the audio channel and/or its parts, taking into account at least a recording codec of the audio channel, available transcoding from the recording codec to a speech recognition engine supported codec, a length of the audio channel, and a language of the audio channel;

applying an audio channel analyzer to the audio channel to characterize audio fidelity, background noise, concurrent speech by multiple speakers, timbre, pitch, and audio distortion of the audio channel; and

using the machine learning engine to select a speech recognition engine that is qualified as capable of transcribing the audio channel and/or its parts or a transcoded version of the audio channel and/or its parts based a combination of at least three of the following characteristics:

audio fidelity;

background noise;

concurrent speech by multiple users;

timbre;

pitch; and

audio distortion of the audio channel.

11. The computer-implemented method of claim 1 , wherein the selecting of a qualified speech recognition engine is based on a combination of least four of the following characteristics:

audio fidelity;

background noise;

concurrent speech by multiple users;

timbre;

pitch; and

audio distortion of the audio channel.

12. The computer-implemented method of claim 11 , further including:

using the multiple speech recognition engines on the audio channel, including using the speech recognition engines sequentially when a first speech recognition engine reports a low confidence score on some or all of its transcription.

13. The computer-implemented method of claim 11 , further including:

using the multiple speech recognition engines on all or separate parts of the audio channel, including using the speech recognition engines when voting on transcription results is used, when different speakers on different tracks of the audio channel, and when different speakers take turns during segments of the audio channel.

14. The computer-implemented method of claim 11 , further including applying the method to auto-punctuation engines.

15. The computer-implemented method of claim 11 , further including applying a silence analyzer to the speech samples and the audio channel prior to submission to parse out silent parts of speech.

16. The computer-implemented method of claim 11 , further including performing testing periodically, including daily, weekly, or monthly, wherein the testing comprises the submitting of hundreds of speech samples to multiple speech recognition engines, and the analyzing of how transcription error rates of the speech recognition engines vary with sound and speech characteristics of the speech samples.