Yes, but the
speech
recognition accuracy varies greatly depending upon a large
number of
factors, including the type of speech (from prepared to spontaneous
speech and conversational speech) and the noise level. So you can
expect very good results when transcribing the speech of an anchor
speaker in a TV or radio news show, but much less good results for
the speech of someone engaged in a very casual conversation.
Yes, the output of the VoxSigma software is an XML file that can be
easily converted into plain punctuated text by discarding additional
information such as word time-codes and word confidence scores.
It depends greatly on the available
language resources for the specific language. It also depends on the type of
speech data you want to process. We are supporting many languages, including
Arabic, Cantonese, Czech, Dutch, English, Finnish, French, German, Greek, Hebrew, Hindi, Hungarian,
Italian, Latvian, Lithuanian, Mandarin, Pashto, Persian, Polish, Portuguese,
Romanian, Russian, Spanish, Swahili, Swedish,
Turkish, Ukrainian and Urdu.
Contact us to get a more precise answer
for the languages you are interested in.
Vocapia Research
LVCSR systems come
with fully trained language models, so the only information you have
to provide to the system is the language being spoken.
If the language is not known, the language can be identified
automatically (among 100 known languages) by using the VoxSigma language recognition
software. A language identification system identifies the language
being spoken from the speech signal.
First you need a speech data set
representative of the targeted data along with a reference transcription. This
data set must large enough to estimate an accuracy which statistically
significant. It is common to use test sets with 3 to 5 hours of speech from
at least 20 speakers. It is common practice to measure
the
word error rate (WER) instead of the
accuracy as it is correlated with the cost of using the system. The WER is
defined as the ratio between the sum of the substitutions, insertions, and
deletion, divided by the total number of word in the reference word. You can
use the NIST sclite software to perform the alignment between the reference
words and hypothesized words and compute the WER and to analyze the errors.
