Pronunciation Discrepancy Between API and Web-Based Audio Generation

[PUJARIGNANASAI]

@smallest-admin We are experiencing inconsistent pronunciation issues when generating audio using the API with voice_id="shweta".
For example, the word "country" is sometimes pronounced correctly, but other times it is truncated to "cont".
However, when generating the same audio via the WEB interface, the pronunciation remains consistently correct.

My code snippet:

client.synthesize(
save_as=os.path.join(output_path,"test.wav"),
add_wav_header=True,
cfg=0,
is_pvc=False,
nfe_step=16,
sample_rate=24000,
save_history=True,
speed=1,
transliterate=False,
text= "Only a very small minority of the people of this country believe in gambling as a substitute for the old philosophy of Benjamin Franklin",
voice_id="shweta"
)

Could you suggest any way to match both outputs.

Audio files: https://drive.google.com/drive/folders/1hZf6LBLh74dIdNSuvYIAlqHraOH68rmC?usp=sharing

[hamees-sayed]

Hey @PUJARIGNANASAI I was able to reproduce the issue on both platform and SDK, It might be due to the seed. Could you try generating it again and let me know if the issue persists?

[PUJARIGNANASAI]

Yeah @hamees-sayed @smallest-admin we tried multiple times, but we are getting the correct pronunciation only 4 out of 10 times.
What should we do to ensure we get it right every time?
By the way, what is that seed, and how can we resolve the seed issue?

I’m uploading 10 re-run output files here:
https://drive.google.com/drive/folders/1INYpE9-xmC9NdAXVf0QISCVkhpHbjtn8?usp=sharing

[jgyasu]

@PUJARIGNANASAI pretty late reply, but I think @hamees-sayed is probably talking about the random seed value, it is something which you set to get a consistent result on every run