jcrocholl

www.nxdom.com is my brainstorming tool for short available domain names. I'm trying to eliminate scrolling and show similar names together. How do you like the condensed display format with several names on each line?

Also, do you prefer the basic or advanced sort options on the left side?

Thanks in advance, I really appreciate your feedback.

I'm an experienced Python/Django developer looking for co-founders in Seattle. I would like to have beer or coffee with experienced web UI designers / developers / business guys. My email address is on my profile.

My favorite startup idea right now is a web service for audio-to-text transcriptions, using Automatic Speech Recognition and Amazon Mechanical Turk. But I'm also interested if you have a different brilliant idea.

I founded browsershots.org in 2005 and sold it in October 2008. This year I finally finished my University degree in Computer Science. I'm from Germany but living in Seattle now.

I want to make online transcription (speech to text) more accurate, faster, and cheaper at the same time. Users can upload MP3 files of interviews or podcasts or any other voice recording, using a web form or an API. Or they can call a phone number, enter their account number with the dial pad, and dictate their message after the beep.

My first step would be to cut the input into 5-minute chunks and use automatic speech recognition (ASR) to generate a rough outline for the transcription. Then each transcription chunk is posted automatically to Amazon Mechanical Turk for proofreading and editing. Turkers can earn points for good work, and this will qualify them for premium tasks which cost more.

The resulting Audio and Text can be used to improve the acoustic models for the speech recognition engine, so the automatic transcripts get better over time, and less work is required for proofreading and editing. It would be possible to train several classes of speaker-independent acoustic models, e.g. adult female speaker with German accent. Languages other than English are possible too.

This service is very similar to castingwords.com but faster and cheaper because it uses self-improving speech recognition technology.

Please let me know what you think. I'm planning to implement a simple prototype in Seattle during the next few weeks. Want to brainstorm with me over beer or coffee? We could be co-founders if we work well together.

I want to solve speech recognition for programmers, so that we can create and edit source code efficiently without using a keyboard or a mouse. This will be great for RSI sufferers (unless voice also fails because of overuse).

Accurate speech recognition is a hard problem because human language is very ambiguous. As far as I know, there's currently no working solution for programming by voice in any programming language. DragonDictate or NaturallySpeaking with custom macros may come close, but it's difficult to set up. I found some related material, but most of it seems to be either academic research or unmaintained software.

Initially, I'm going to focus on speech recognition for Python. I know there's probably a bigger market for C++ or Java programmers, but Python code is more similar to human speech and is rather concise, using fewer lines of code than Java for the same task. Python has a large standard library which we can pre-parse and digest to reduce ambiguity during recognition.

Python's interactive interpreter with speech recognition and voice output would make an awesome demo. You could say "three times five" and the computer would respond with "fifteen". Or you could say "from time import localtime (pause) call localtime without parameters slice the first three elements" and the system would say "two thousand nine eight seventeen".

The entire speech recognition software could be free open source, maybe based on CMU Sphinx-4 (which is in Java). The business model could revolve around a web service that lets people upload their utterances (snippets of recorded speech) during or after their programming session. We can use these files to improve the recognition engine and train the speaker-independent acoustic model. So the recognition would get better over time, but speaker-independent models only work for some "standard" pronunciation without too much accent.

For a small fee (e.g. $49) users could download their personal acoustic model for improved accuracy, which would be generated from the voice snippets that they have uploaded. The model training process needs several minutes or even hours of CPU, but an email could be sent to the user when the model is ready for download. When the software improves over time and they have recorded more utterances, they can pay another fee and generate an even better model.

If I can get speech recognition for Python code to work, maybe SQL or bash could work too (both support auto-completion, which can be useful for reducing ambiguity).

Please let me know what you think. I'm planning to implement a simple demo in Seattle during the next few weeks. Want to brainstorm with me over beer or coffee? We could be co-founders if we work well together.