Logs for speechsc@ietf.xmpp.org, 2004-08-03[14:53:42] --- eburger has joined [14:58:21] --- eburger has left: Replaced by new connection [14:58:22] --- eburger has joined [14:58:22] --- eburger has left [15:03:41] --- eburger has joined [15:04:53] About to start... [15:05:04] Slides at http://flyingfox.snowshore.com/i-d/speechsc [15:05:21] Sarvi: Status on MRCPv2 [15:08:11] Discussion on: Text phrases should be allowed as required phrases, in addition to digit sequences. VERIFY-FROM-BUFFER & RECOGNIZE [15:08:20] (from Claudia Daboul) [15:10:14] If verification buffer filling when VERIFY-FROM-BUFFER received; then server waits for buffer to fill. Silently waits. [15:10:27] Do get In Progress return code [15:11:05] Klaus: Don't like requiring buffering mechanism (in general) to be required. Was optional. [15:11:32] Sarvi: only will issue VERIFY-FROM-BUFFER for old utterances, anyway. I.e., only if you do buffering. [15:12:40] Next [15:12:52] Accept-Threshold and Adapt-Threshold [15:12:58] (also form Claudia) [15:13:29] Need to clarify behavior of VERIFY-FROM-BUFFER while buffering (last item) [15:13:31] new: [15:13:47] Security Level: What level of verification score counts for success? [15:14:32] Accept threshold - what counts. Proposal: Adapt Threshold - higher level of acceptance to adapt to current voice state. [15:14:54] Dave O: why isn't this a pure client thing? Just report result and let client decide what to do? [15:17:36] --- eburger has left: Replaced by new connection [15:17:36] --- eburger has joined [15:17:36] --- eburger has left [15:17:42] --- eburger has joined [15:18:27] Dan B.: score is vendor dependent; use normal scale from client [15:18:43] Dave O.: Calling anything "Security Level" is big red flag. [15:19:40] Dave O.: Propose: "Server Score" and "Confidence Level" as separate elements [15:20:12] Dan B.: Confidence Level is "High, Medium, Low", not a number [15:21:22] Sarvi - can we map from numbers to high/medium/low values? [15:21:34] Eric: Why not have server return High/Medium/Low? [15:21:41] Dan: Could work. [15:23:10] --- moritanaotaka has joined [15:24:00] Dave: Dan brings up issue: does score represent error rates, or to have protocol say something like "I can live with false positive rate of X and false negative rate of Y"? [15:25:04] Dave - confidence is 1-dimensional metric, but client cares about 2-dimensional problem [15:25:34] client cares about Yes/No based on client's meaningful requirements for true/false positive/negative needs [15:26:06] Take issue to list. [15:30:45] --- moritanaotaka has left [15:31:38] Voice Print Group: [15:31:46] --- naotaka has joined [15:32:01] partially a DB problem, partially an organization problem. [15:32:08] Dave: Won't do DB stuff in speechsc [15:33:25] Dan - is there a difference between a voice print and a voice print group from a protocol perspective? [15:33:51] Sarvi - don't have hard requirement to map input voice print to result voice print [15:34:03] Klaus - how do you know what is in the group? [15:34:23] Sarvi - already opaque object to protocol; vendor specific [15:35:51] Do we want to introduce concept of Voice Print Groups (but not the management part)? [15:37:15] Dave - Is Group an equivalence class (any match is a match) or is Group an optimization to identify a given person? [15:39:58] Group really is pass-by-reference [15:42:19] Trade-off of complexity in client versus server. Take to list. [15:42:31] Voice-Print adaptation in multi-verification result format. [15:42:48] Only one voice print: easy to decide if to adapt print. [15:43:09] Multiple voice prints: do we want to adapt best match? All in set? Nothing? [15:43:24] Dan (and consensus) - not interesting to update all in set [15:43:53] Any objections to banning adaptation for multi-verification results? [15:44:08] Klaus: Yes, still should be OK to adapt best result. [15:44:34] Dan: Single verification: can do adaptation "on the fly" as utterances come in. [15:44:54] Multi verification: not possible to do adaptation on the fly. [15:45:01] Klaus: this is why we have training from buffer [15:45:22] Dan: also, what if you get two "good" results? [15:46:02] Dave - Verification is idempotent; adaptation is state changing [15:46:18] single verification, straightforward what state gets changed [15:46:38] Multi-verification, which gets state changed? [15:46:55] [Back to issue of two "winners" of verification] [15:50:04] Dave - we will take speech marker time stamps to list [15:52:13] --- eburger has left [16:01:56] --- naotaka has left: Disconnected