Hey folks, data_shipper_joe here! I usually hang out talking about moving data from point A to point B, but lately I've been playing with Resemble AI's API for a side project generating synthetic voiceovers for training data narration. Hit a weird snag and wanted to see if anyone else has experienced this.
I'm using the Python SDK to generate speech and download the WAV files. The generation works great, but when I concatenate longer scripts (using the `ssml` method with multiple sentences), the final downloaded files have these subtle, but audible, clicks or glitches at the boundaries between the segments Resemble seems to generate internally. It's almost like a tiny pop or a stutter where one phrase ends and the next begins. This doesn't happen with very short, single-sentence clips.
Here's the basic flow I'm using:
```python
import resemble
# ... setup and script split into paragraphs ...
response = resemble.resemble.create_voice_clip(
project_uuid=project_uuid,
voice_uuid=voice_uuid,
body=ssml_content, # e.g., "Paragraph one. Paragraph two."
precision="high"
)
# Download the WAV from the response URL
```
If I generate each paragraph as a separate clip and stitch them together locally with pydub, the issue is gone. But that's extra API calls and feels like a workaround. It seems like the artifact is introduced during Resemble's own internal synthesis stitching when handling longer SSML.
Has anyone else run into this? Wondering if it's a known issue with the audio encoding/concatenation on their end, or if there's a specific `precision` or output format setting that might mitigate it. I'm on a `premium` voice model.
Appreciate any tips—trying to get this pipeline as clean as the data pipelines I usually build!
ship it