Known limitations

Issue: Interim results for previous-generation models

27 April 2021: When you use previous-generation models with the WebSocket interface, if you request both speaker labels and interim results, the speaker labels response includes an extra object. The final results duplicate the object for the last speaker label. Instead of appearing once with "final": true, the object appears twice: once with "final": false and once with "final": true.

For example, a WebSocket request that includes both speaker labels and interim results sends a start message like the following:

var message = {
  action: 'start',
  speaker_labels: true,
  interim_results: true
};
websocket.send(JSON.stringify(message));

To indicate final results for speaker labels, the service is supposed to return a response of the following form:

{
  "speaker_labels": [
    . . .
    {
      "from": 1.01,
      "to": 1.75,
      "speaker": 0,
      "confidence": 0.75,
      "final": false
    },
    {
      "from": 1.76,
      "to": 2.50,
      "speaker": 1,
      "confidence": 0.80,
      "final": true
    }
  ]
}

Instead, the service returns final results like the following. Note that the object for the last speaker label, which begins at 1.76 and ends at 2.50, is returned twice. The "final": true indication is associated with the second instance of the label.

{
  "speaker_labels": [
    . . .
    {
      "from": 1.01,
      "to": 1.75,
      "speaker": 0,
      "confidence": 0.75,
      "final": false
    },
    {
      "from": 1.76,
      "to": 2.50,
      "speaker": 1,
      "confidence": 0.80,
      "final": false
    }
  ]
}{
  "from": 1.76,
  "to": 2.50,
  "speaker": 1,
  "confidence": 0.80,
  "final": true
}

For more information about using speaker labels with interim results, see Requesting interim results for speaker labels.

Issue: Supported features for speaker labels is always true

6 August 2020: The GET /v1/models and GET /v1/models/{model_id} methods list information about language models. Under supported_features, the speaker_labels field indicates whether you can use the speaker_labels parameter with a model. At this time, the field returns true for all models.

However, speaker labels are supported as beta functionality only for US English, Australian English, German, Japanese, Korean, and Spanish (both broadband and narrowband models) and UK English (narrowband model only). Speaker labels are not supported for any other models. Do not rely on the field to identify which models support speaker labels.

For more information about speaker labels and supported models, see Speaker labels.

Issue: The progress field for custom models is not continuous

The following methods include a progress field in the information that they return for a custom model:

• GET /v1/customizations
• GET /v1/customizations/{customization_id}
• GET /v1/acoustic_customizations
• GET /v1/acoustic_customizations/{customization_id}

The field returns the progress for a training or upgrade operation. Currently, the value of the field is 100 if the status is available; otherwise, it is 0.

When you monitor the training or upgrading of a custom model, poll the value of the status field, not the value of the progress field. If the operation fails for any reason, the value of the status field changes to reflect the failure; the value of the progress field remains 0.