With Google Cloud Speech-to-text, why do I get different results for the same audio file, depending on which bucket do I put it into? - google-cloud-speech

Question

I am trying to use Google Cloud Speech-to-text, using the client libraries, from a node.js environment, and I see something I don't understand: I get a different result for the same example audio file, and the same configuration, depending on whether I am using it from the original sample bucket, or from my own bucket.
There are the requests and responses:
The baseline is Google's own test data file, available here: https://storage.googleapis.com/cloud-samples-tests/speech/brooklyn.flac
Request:
{
"config": {
"encoding": "FLAC",
"languageCode": "en-US",
"sampleRateHertz": 16000,
"enableAutomaticPunctuation": true
},
"audio": {
"uri": "gs://cloud-samples-tests/speech/brooklyn.flac"
}
}
Response:
{
"results": [
{
"alternatives": [
{
"transcript": "How old is the Brooklyn Bridge?",
"confidence": 0.9831430315971375
}
]
}
]
}
So far, so good. But, if I download this audio file, re-upload it to my own bucket, and do the same, then:
Request:
{
"config": {
"encoding": "FLAC",
"languageCode": "en-US",
"sampleRateHertz": 16000,
"enableAutomaticPunctuation": true
},
"audio": {
"uri": "gs://goe-transcript-creation/brooklyn.flac"
}
}
Response:
{
"results": [
{
"alternatives": [
{
"transcript": "how old is",
"confidence": 0.8902621865272522
}
]
}
]
}
As you can see this is the same request. The re-uploaded audio data is here: https://storage.googleapis.com/goe-transcript-creation/brooklyn.flac
This the exact same file as in the first example... not a bit of difference.
Still, the results are different; I only get half of the sentence.
What am I missing here? Thanks.
Update 1:
The same thing happens with the CLI tool, too:
$ gcloud ml speech recognize gs://cloud-samples-tests/speech/brooklyn.flac --language-code=en-US
{
"results": [
{
"alternatives": [
{
"confidence": 0.98314303,
"transcript": "how old is the Brooklyn Bridge"
}
]
}
]
}
$ gcloud ml speech recognize gs://goe-transcript-creation/brooklyn.flac --language-code=en-US
ERROR: (gcloud.ml.speech.recognize) INVALID_ARGUMENT: Invalid recognition 'config': bad encoding..
$ gcloud ml speech recognize gs://goe-transcript-creation/brooklyn.flac --language-code=en-US --encoding=FLAC
ERROR: (gcloud.ml.speech.recognize) INVALID_ARGUMENT: Invalid recognition 'config': bad sample rate hertz.
$ gcloud ml speech recognize gs://goe-transcript-creation/brooklyn.flac --language-code=en-US --encoding=FLAC --sample-rate=16000
{
"results": [
{
"alternatives": [
{
"confidence": 0.8902483,
"transcript": "how old is"
}
]
}
]
}
It's also interesting that when pulling the audio from the other bucket, I need to specify encoding and sample rate, otherwise it doesn't work... but it's not necessary when I am using the original test bucket.
Update 2:
If I don't use Google Cloud Storage, but upload the data directly in the speech-to-text request, it works as intended:
$ gcloud ml speech recognize brooklyn.flac --language-code=en-US
{
"results": [
{
"alternatives": [
{
"confidence": 0.98314303,
"transcript": "how old is the Brooklyn Bridge"
}
]
}
]
}
So the problem doesn't seems to be with the recognition itself, but accessing the audio data. The obvious guess would be that maybe it's the fault of the uploading, and the data is somehow corrupted along the way?
We can verify that by pulling the data from the cloud, and comparing with the original. It doesn't seem to be broken.
So maybe it's a problem when the S-T-T service is accessing the storage service? But why with one bucket only? Or is it some kind of file metadata problem?