Hey folks, been diving into the audio-to-text API space for a side project. I'm scraping and analyzing earnings call transcripts and financial news audio clips, trying to extract sentiment and key entities. My usual go-to has been Google's Speech-to-Text (with its NLP features), but I've been hearing a lot about Cartesia's real-time capabilities and their focus on low-latency streaming.
I'm specifically working with recent data (last 30 days) from a few financial news podcasts and YouTube streams. My main metrics are:
* **Accuracy:** Especially on financial jargon, company names, and numbers.
* **Latency:** For a near-real-time dashboard I'm prototyping.
* **Cost per hour:** Since I'm processing a decent volume.
I set up a quick Python script to send the same audio chunks to both APIs and compare the JSON outputs. Cartesia's API is pretty clean:
```python
# Cartesia snippet
response = requests.post(
"https://api.cartesia.ai/tts/...",
headers={"Authorization": f"Bearer {api_key}"},
json={"model_id": "sonic-english", "input": audio_bytes}
)
```
Has anyone run a more thorough, apples-to-apples benchmark on **recent news data**? I'm curious about:
- Word Error Rate (WER) on fast-talking news anchors.
- How each handles overlapping speakers in panel discussions.
- The real-world difference in `endpoint_detection` for streaming.
My initial, very rough tests show Cartesia might be faster on latency, but I'm not fully convinced on accuracy for niche terminology compared to Google's `video` model. Also, Google's natural language processing (like entity extraction) feels more baked-in, whereas with Cartesia I'd probably pipe the text to another model.
Would love to see if anyone has done a structured comparison or has a GitHub repo to share. Pricing models are also a factor—Cartesia's per-second billing could be interesting for short clips.
--diver
Data is the new oil - but it's usually crude.