google speech recognition weak date transcription - google-cloud-platform

Question

I am using currently google asr/tts with german speech model (de-DE) and experience wrong results in certain usage scenarios of date extractions.
I am really keen on to know whether others might have similar results.
Let me give you some examples:
I am saying:
"der 1.10.1905" -> "1. 10195 11095"
"9.5.78" -> 90587
"22.11.98" -> 22.11 89
BUT:
"22. November 98" -> "22 November 98"
When I fully qualify the month word it works all fine.
I also checked the proposed way to optimize by using hints, without any improvement
"speechContexts": [
{
"phrases": [
"$FULLDATE"
]
}
]
Is this something one has to accept or is there anything that I can try on top?
Cheers Andre
UPDATE:
According to https://issuetracker.google.com/issues/186559805 the problem should be fixed.
But
I could not verify it, maybe somebody else can ??!
I tried it with the following configuration, but it did not improve at all. Maybe overlooked something ?!
Here is my configuration. I performed the request with the beta client:
const {SpeechClient} = require('#google-cloud/speech').v1p1beta1;
const googleRequest = {
config: {
encoding: "MULAW",
sampleRateHertz: 8000,
model: "command_and_search",
languageCode: "de-DE",
speechContexts: [
{
"phrases": [
"$OOV_CLASS_FULLDATE"
]
}
]
},
singleUtterance: true,
interimResults: true
};
I also raised a new ticket to google developers.
https://issuetracker.google.com/issues/243294056

Answer 1

I have the same results as you have when using your example. But it does transcribe correctly if the speaker says ein tausend neun hundert acht und neunzig instead.
It seems that, for some reason, in German, the API doesn't recognize when the speaker uses hundreds to express thousands (I.E. nineteen hundred ninety eight instead of one thousand nine hundred ninety eight).
I don't know German and can't confirm if this way of speaking numbers/dates is official but if you are sure that in German, it is officially accepted, then I suggest that you create an issue on issue-tracker with some references to confirm it.