• 

  Olewave Speech Data Processing Pipeline

• 

  The pipeline is the only one in the market can reliably generate word timestamps even for overlapped speech conversations.

Why Voice AI Needs Word Timestamps?

For Business Analytics:

• Keywords Spotting:
  For example, when did the CFO mention ‘spending’ in the earnings call?

For Training Voice AI Models:

• Turn-Taking Model:
  Needs accurate speaker turn data.
• Text-to-Speech Model:
  Requires data on word or phone duration.
• Meeting Summarization Model:
  Needs data on ‘who spoke when’.

Why OpenAI’s Whisper is Not Good at Generating Word Timestamps?

• Not Trained with Timestamp-Related Criteria:
  Whisper models are primarily optimized to predict the most probable sequence of wordpieces or tokens, rather than ensuring precise temporal alignment between the audio and the corresponding text. The activations in attention weights or the spike triggers in CTC (Connectionist Temporal Classification) do not inherently guarantee accurate timestamp alignment with the actual speech.

• Postprocessing Whisper’s Results is Not Recommended:
  Some open-source tools have been developed for ad-hoc postprocessing, but they do not follow industrial standards and are not trained on proper datasets. While some efforts are on the right track, they often use unfit open-source datasets instead of appropriate ones for training the AI model.

Olewave’s Speech Data Processing Pipeline

For more details, refer to the figure provided.

If you want to know more about our services and datasets, feel free to send an email to info@olewave.com and see how we can assist you!

Built With

• olewave-techstack