feat: SpeechToDocument

[ZanSara]

Proposed changes:

• Introduces SpeechToDocument, a node that takes in input audio files and outputs a Document.
  • What works:
    • Can deal with arbitrarily long audio files by chunking them internally into fragments
    • The speed of transcription is faster than the audio by about 2x, so in theory it could be used with streaming inputs (but in practice it's very hard now due to how the rest of Haystack works 😄 ).
    • The audio is aligned with its transcription at the word level using aeneas ("traditional" forced alignment method, so very fast)
  • Still to do, probably in the next PR:
    • Support input files other than .wav 😅 (easy)
    • Denoising the input audio (easy to remove most of it, very hard to make it perfect)
    • Adding punctuation or test with models that predict punctuation too (to investigate)
    • Fragmenting the input on voice pauses (now broken down in arbitrarily long chunks) (to investigate)
    • Silence detection (should be relatively easy)
    • Run a spellchecker on the output to improve the transcription quality (should be relatively easy)
• Modifies AnswerToSpeech to check if the source document contains alignment data: if it does, the new AnswerToSpeech will extract the audio from the original source instead of generating it.
• Introduces a new primitive, AudioAlignment, that contains alignment data
• Modifies the SpeechDocument/SpeechAnswer primitives to accommodate for optional alignment data
• Modifies Span by overriding its in statement. Now assert 10 in Span(5, 15) returns True 😁

To Dos:

• Tutorial
• Tests

[ZanSara]

These nodes will be added to Haystack as external nodes. Closing.