dmkeeler6

My audio topic was the first three-minute run-through of my podcast focusing on how house systems impact school culture. The tool I used to create the transcript was Otter.ai. I was surprised how well the AI translated the audio. I used both their and they’re in the audio and it transcribed the correct homophone. It did well with proper nouns and even capitalized them. It messed up a couple of times with filler words. It did try to transcribe them but occasionally guessed wrong. It also tried to transcribe sounds I made as words. At one point, I made an inarticulate filler noise at the beginning of a sentence that AI interpreted as the word “so”. The biggest thing I noticed was it had difficulty at times when I did not articulate well or when I spoke a little too fast. During the introduction of my audio, I said “Whether you teach math, science, English or even art you’re sure to pick up usable classroom tips and tricks…” I was speaking quickly in this section and the AI interpreted “art you’re sure to pick up” as “articulable”. It seems like when AI cannot fully understand what a speaker is saying, it tries to guess using context and often gets it wrong. I noticed that the AI transcript is just one big wall of text. It does not sort out the audio into sections or paragraphs. The AI does not always parse the audio into proper sentences or take cadence or sentence flow into account. A pause that may be used by the speaker in audio does not show up in the transcript, which means the nature of the audio can sometimes get lost to a reader. I think when audio is short and has one speaker who is speaking clearly and concisely, AI can be trusted to transcribe the audio. For longer audio where segmenting is important and the speaker may be addressing multiple topics, it is important to hand-correct to ensure the structure of the transcript (paragraphs, sentence structure) properly communicates the message to the reader. I also think it would be important to check when there are multiple speakers, especially when they have similar voices and cadence because I suspect AI would do a poor job differentiating between similar voices.