3 ms·

Even Amazon, with all the server capacity in the world, is not going to be able to store and process every sound heard by every Alexa in real time.

by paxy 9y ago

throwaway2048 9y ago
no, but its entirely feasable to store all spoken word transcripts.
pdkl95 9y ago
> is not going to be able to store Voice codecs like AMR[1] (or GSM-HR) only need 2-3MB/hr. A trivial noise gate[2] that cuts out silence should allow a recording duty cycle under ~1%. Amazon can easily store all the voice recording they want. > and process ... in real time Processing isn't that expensive once you have a trained classifier. However, realtime processing often isn't necessary. [1] https://en.wikipedia.org/wiki/Adaptive_Multi-Rate_audio_codec https://en.wikipedia.org/wiki/Adaptive_Multi-Rate_audio_code... [2] https://en.wikipedia.org/wiki/Noise_gate https://en.wikipedia.org/wiki/Noise_gate
- rootw0rm 9y ago
  Excellent response.
- zeta0134 9y ago
  I don't doubt that what you say is correct, but I think the author's point is that even with those optimizations, the sheer amount of data from [Every Alexa in the world] * [Every second of every day] is too much for any remote service to process. The wake word needs to be handled locally so that the remote service is only involved when needed. The amount of data for a single recording can be quite small, but multiplied by a large population it still adds up in a hurry.
  - pdkl95 9y ago
    > still adds up in a hurry Never mind the wake word; a noise gate (which is a trivial computation locally in either hardware[1] or software) means we are not talking about [every second of every day]. I'm sure Amazon/whomever can do better than a simple noise gate, too. Using GSM-HR which records voice at 5.6 kbit/s (a bit over 700 bytes/s). 5.6 * 60 * 60 / 1024 / 8 = 2.4609375 MB/hr Assuming a gate that only records 1% of the time, this is 2.4609375 * 0.01 = 0.024609375 MB/hr = 25.2 kB/hr = 604.8 kB/day 604.8 * 365.25 / 1024 = 215.72578125 MB/year Assuming 40M devices[2], storing all the voice said around these devices for a year is only ~8 PB (8229.28 TB). Renting the necessary storage from Amazon AWS, from a brief look at their current prices[3] costs at most $200,000/month ("standard storage"). Yes, the storage costs add up over time, but the amount of data needed to record all the voice said around every Alexa in the world is trivially within Amazon's storage capability. [1] voice activated tape recorders have been around for decades [2] https://techcrunch.com/2018/01/12/39-million-americans-now-own-a-smart-speaker-report-claims/ https://techcrunch.com/2018/01/12/39-million-americans-now-o... [3] https://aws.amazon.com/s3/pricing/ https://aws.amazon.com/s3/pricing/

Even Amazon, with all the server capacity in the world, is not going to be able to store and process every sound heard by every Alexa in real time.

by paxy 9y ago

Even Amazon, with all the server capacity in the world, is not going to be able to store and process every sound heard by every Alexa in real time.

throwaway2048 9y ago

no, but its entirely feasable to store all spoken word transcripts.

pdkl95 9y ago

> is not going to be able to store Voice codecs like AMR[1] (or GSM-HR) only need 2-3MB/hr. A trivial noise gate[2] that cuts out silence should allow a recording duty cycle under ~1%. Amazon can easily store all the voice recording they want. > and process ... in real time Processing isn't that expensive once you have a trained classifier. However, realtime processing often isn't necessary. [1] https://en.wikipedia.org/wiki/Adaptive_Multi-Rate_audio_codec https://en.wikipedia.org/wiki/Adaptive_Multi-Rate_audio_code... [2] https://en.wikipedia.org/wiki/Noise_gate https://en.wikipedia.org/wiki/Noise_gate

rootw0rm 9y ago
Excellent response.
zeta0134 9y ago
I don't doubt that what you say is correct, but I think the author's point is that even with those optimizations, the sheer amount of data from [Every Alexa in the world] * [Every second of every day] is too much for any remote service to process. The wake word needs to be handled locally so that the remote service is only involved when needed. The amount of data for a single recording can be quite small, but multiplied by a large population it still adds up in a hurry.
- pdkl95 9y ago
  > still adds up in a hurry Never mind the wake word; a noise gate (which is a trivial computation locally in either hardware[1] or software) means we are not talking about [every second of every day]. I'm sure Amazon/whomever can do better than a simple noise gate, too. Using GSM-HR which records voice at 5.6 kbit/s (a bit over 700 bytes/s). 5.6 * 60 * 60 / 1024 / 8 = 2.4609375 MB/hr Assuming a gate that only records 1% of the time, this is 2.4609375 * 0.01 = 0.024609375 MB/hr = 25.2 kB/hr = 604.8 kB/day 604.8 * 365.25 / 1024 = 215.72578125 MB/year Assuming 40M devices[2], storing all the voice said around these devices for a year is only ~8 PB (8229.28 TB). Renting the necessary storage from Amazon AWS, from a brief look at their current prices[3] costs at most $200,000/month ("standard storage"). Yes, the storage costs add up over time, but the amount of data needed to record all the voice said around every Alexa in the world is trivially within Amazon's storage capability. [1] voice activated tape recorders have been around for decades [2] https://techcrunch.com/2018/01/12/39-million-americans-now-own-a-smart-speaker-report-claims/ https://techcrunch.com/2018/01/12/39-million-americans-now-o... [3] https://aws.amazon.com/s3/pricing/ https://aws.amazon.com/s3/pricing/