Cell phones and phonetics
Doing some doc-reading at work today. My focus might be switching briefly from the NTCIP work that I was ramping up for to some light GPRS work. Well, since I'm not really familiar with the topic -- like, at all -- I started ramping up on the technologies by browsing Wikipedia. I love Wikipedia. GPRS led me to GSM. GSM is apparently rooted in LPC.
This is where things get interesting. See, according to Wikipedia, LPC starts with the mathematical analysis tools of maximum likelihood estimation and the principle of maximum entropy, and then it applies them to phoneme recognition. If I understand correctly, the distinctive properties of these phonemes themselves are then encoded and sent over the medium (Internet for VoIP, cell network for cell modems) to the other side of the conversation, where they're reconstructed into an actual waveform and spat out a speaker.
Now, here's my question... Does LPC actually encode phonemes, or does it encode phones? See, when we talk about the sound of the letter "p," for instance, we're actually talking about a whole range of actual sounds. We think of it as one sound, but that's because English doesn't distinguish between, for instance, an aspirated p and an unaspirated p. You probably won't even hear the difference between "pin" (aspirated p) and "spin" (unaspirated p) unless you really listen carefully for it. Aspirated p and unaspirated p are different phones (they look different on a waveform analyzer), but they're the same phoneme in English because we use them interchangeably. That "in English" qualifier is the key part, though. Other languages (e.g., Chinese, Hindi, ancient Greek) make a distinction between the aspirated and unaspirated versions of some consonants. They're different phonemes in those languages.
So if LPC encodes phonemes, then which language's phonemes does it encode? And if it encodes, for instance, English phonemes, then does that have an adverse effect on the quality of speech reproduction in languages with different phoneme sets, such as those that distinguish between aspirated and unaspirated consonants?
(LJ Spellchecker Genius of the Day: VoIP -> Poop)
This is where things get interesting. See, according to Wikipedia, LPC starts with the mathematical analysis tools of maximum likelihood estimation and the principle of maximum entropy, and then it applies them to phoneme recognition. If I understand correctly, the distinctive properties of these phonemes themselves are then encoded and sent over the medium (Internet for VoIP, cell network for cell modems) to the other side of the conversation, where they're reconstructed into an actual waveform and spat out a speaker.
Now, here's my question... Does LPC actually encode phonemes, or does it encode phones? See, when we talk about the sound of the letter "p," for instance, we're actually talking about a whole range of actual sounds. We think of it as one sound, but that's because English doesn't distinguish between, for instance, an aspirated p and an unaspirated p. You probably won't even hear the difference between "pin" (aspirated p) and "spin" (unaspirated p) unless you really listen carefully for it. Aspirated p and unaspirated p are different phones (they look different on a waveform analyzer), but they're the same phoneme in English because we use them interchangeably. That "in English" qualifier is the key part, though. Other languages (e.g., Chinese, Hindi, ancient Greek) make a distinction between the aspirated and unaspirated versions of some consonants. They're different phonemes in those languages.
So if LPC encodes phonemes, then which language's phonemes does it encode? And if it encodes, for instance, English phonemes, then does that have an adverse effect on the quality of speech reproduction in languages with different phoneme sets, such as those that distinguish between aspirated and unaspirated consonants?
(LJ Spellchecker Genius of the Day: VoIP -> Poop)