Improve API response structure for timestamps and word_confidence within STT SpeechRecognitionAlternative model

As part of the response from making  a POST to the v1/recognize endpoint in the Speech to Text service, the user receives an array of "alternatives". Within these "alternatives" objects, there are two arrays called "word_confidence" and "transcript". Below is an example of this piece of the response:

"alternatives": [
  {
    "transcript": "thunderstorms could produce",
    "confidence": 0.994,
    "word_confidence": [
      [
        "thunderstorms",
        1
      ],
      [
        "could",
        1
      ],
      [
        "produce",
        1
      ],
    ],
    "timestamps": [
      [
        "thunderstorms",
        1.49,
        2.32
      ],
      [
        "could",
        2.32,
        2.54
      ],
      [
        "produce",
        2.54,
        3.01
      ],
    ]
  }
]

The problem with this response is that both "word_confidence" and "timestamps" are arrays of arrays, even though the internal arrays have specific attributes at particular indices. As is, this response lends itself more to "word_confidence" and "timestamps" being arrays of objects like so:

"alternatives": [
  {
    "transcript": "thunderstorms could produce",
    "confidence": 0.994,
    "word_confidence": [
      {
        "thunderstorms",
        1
      },
      {
        "could",
        1
      },
      {
        "produce",
        1
      },
    ],
    "timestamps": [
      {
        "thunderstorms",
        1.49,
        2.32
      },
      {
        "could",
        2.32,
        2.54
      },
      {
        "produce",
        2.54,
        3.01
      },
    ]
  }
]

Added motivation for the change is that this is currently causing a problem with the API specification and, in turn, efforts to automatically generate code/documentation based on the API.

Currently, we are specifying the various Watson APIs based on the OpenAPI version 2.0 specification. These can be found here. The way the response is actually structured, with a nested array of mixed types, cannot be properly specified according to the specification, resulting in manual work to tweak code that may be generated from the Speech to Text API spec.

Overall, making the proposed change to the v1/recognize response in the Speech to Text service would both make more sense based on the structure, but would also make it easier to document and help further the effort to streamline API changes across all relevant tools.