Tried using Resemble AI's API to generate synthetic speech for an ASR model training experiment. Goal: augment a small dataset of command phrases.
Ran a standard test: generated 1000 audio clips from text prompts, using different voices and emotional tones. Processed them through a Whisper baseline model and a custom Kaldi setup.
Results:
* **Audio Quality:** High. Natural prosody, clear articulation.
* **Consistency:** Excellent. Same voice parameters produced near-identical vocal characteristics across clips.
* **Processing Time:** Average. ~2.1 seconds per clip via API batch.
* **Impact on WER:** Mixed. Added synthetic data reduced WER on in-domain test set by ~8%. Generalization to noisy real-world data showed negligible improvement (<1%).
* **Cost:** Prohibitive for large-scale runs. Generating 100 hours of speech for serious training would be expensive compared to other synthesis or perturbation methods.
Main issue for ASR training is lack of controlled noise and acoustic variation. The audio is too clean. You'd need to layer in background noise, reverberation, etc., post-generation, which adds steps.
Config used for the primary voice generation:
```python
{
"voice": "natalie",
"emotion": "neutral",
"sample_rate": 22050,
"text": "",
"output_format": "wav"
}
```
Has anyone else benchmarked it for this use case? Curious about cost/accuracy trade-offs for larger datasets.
- bench_beast
Benchmarks don't lie.
Your cost point is critical. For 100 hours of data, you're looking at thousands even with bulk rates. At that price, you could commission a small professional voice dataset.
The cleanliness is the real problem for ASR. You need channel effects and noise that match your deployment environment. Slapping random noise on top in post rarely creates the coherent acoustic models you need. The synthesis becomes a bottleneck in the augmentation pipeline.
Have you compared the WER impact against simple speed/volume perturbation of your original data? For command phrases, I've found aggressive time-stretching and controlled additive noise often outperforms expensive synthetic voices on out-of-domain tests.
Agreed on cost. For that budget, you could hire a few contractors on Upwork to record clean, scripted audio in controlled conditions - you'd own the data outright.
"coherent acoustic models" is the key issue. Synthetic data often fails to capture real microphone artifacts and room impulse responses. If your deployment is on mobile devices in noisy environments, you're training on a sanitized version of reality.
I've seen speed/pitch perturbation + domain-specific noise injection (think cafe ambience, car interior) beat generic synthetic speech on WER for wake-word detection. The synthetic data gave a false sense of progress on clean test sets.
Trust, but verify