1.

In 2010, Google couldn't see inside a video — so SpeakerText built a $2-per-minute assembly line where Carnegie Mellon's speech software does the rough cut and Amazon Mechanical Turk workers, editing five-second chunks, do the rest. The bet: make video searchable before the platforms do it themselves.

2.

Web video is invisible to the only audience that matters

• Google indexes titles, descriptions, and meta tags — everything except the actual words spoken inside a video
• Transcription solves this, but professional services charge $3 to $5 per minute, pricing out all but the largest publishers
• Every untranscribed video is a dark room search engines walk past

3.

Sphinx-4 writes the rough draft; Mechanical Turk humans rewrite it

• Carnegie Mellon's open-source Sphinx-4 engine generates an automated first pass — fast but error-prone on unconstrained speech
• SpeakerText slices each video into 5-to-8-second chunks and distributes them to Mechanical Turk workers who correct text and punctuation
• Workers are ranked by performance history, routing harder chunks to proven transcribers
• NLP software reassembles the corrected fragments, detects sentence boundaries, adds timestamps, and generates SEO meta tags

4.

The SpeakerBar turns a transcript into a navigation tool

• Transcribed text appears in a collapsible window beneath the video player — every word visible to search crawlers
• Each sentence is time-stamped and clickable; readers jump straight to the moment they care about
• Copy-pasting any transcript excerpt auto-embeds a deep link back to that exact video timestamp
• Compatible with YouTube, Brightcove, and Blip.tv, plus a WordPress plug-in for self-hosted sites

5.

$2 per minute undercuts the industry floor by 33 to 60 percent

• SpeakerText charges $20 per month for the SpeakerBar plus $2 per minute for transcription
• Competing services start at $3 to $5 per minute and deliver a raw text file — no player integration, no timestamps, no deep links
• The hybrid pipeline's cost advantage widens as feedback loops improve automation and reduce the human correction load over time

6.

The entire value proposition rests on a word: "should"

• SpeakerText's own pitch says transcripts "should help drive more search traffic" — no A/B tests, no publisher case studies, no traffic data cited anywhere in the announcement
• Sphinx-4 in 2010 carried word-error rates of 30 to 50 percent on open-domain speech, meaning Mechanical Turk workers correcting decontextualized five-second audio clips are performing wholesale reconstruction, not light editing — and no accuracy metric for the finished product is provided
• If the SEO premise is unproven and the transcript quality is unverified, SpeakerText is a clever engineering exercise solving a problem that may not exist at any price — and the $2-per-minute advantage is meaningless if the output isn't reliable enough to publish

7.

Sphinx-4 plus Mechanical Turk in 2010 is GPT plus RLHF raters in 2024

• Both the original argument and the counter-argument treat SpeakerText as a transcription company — they share the assumption that its value lives or dies with this specific product in this specific market
• Reframe: the real innovation is the architecture — structured human micro-tasks repairing machine-generated output at scale, with worker ranking and feedback loops. This is the exact blueprint later adopted for data labeling, AI fine-tuning, and reinforcement learning from human feedback
• What's actually at stake is not whether one angel-funded startup survives the day YouTube adds auto-captions — it is whether the human-in-the-loop assembly line is the correct paradigm for every AI task where pure automation falls short. SpeakerText built the prototype; a decade later, the entire AI industry runs on it