Add hosted inference section

README.md

@@ -74,6 +74,24 @@ for word in hypotheses[0].timestamp["word"]:
     print(word["start"], word["end"], word["word"])
 ```
 
+## Hosted inference
+
+If you would rather not run the weights yourself, Pianissimo is served by
+[Berget AI](https://berget.ai) on Swedish infrastructure. The realtime endpoint is compatible
+with the OpenAI Realtime API, so an existing client needs a new base URL and key.
+
+```
+wss://api.berget.ai/v1/realtime?intent=transcription
+model: klang/pianissimo
+```
+
+Audio goes in as PCM16 mono and transcript deltas come back while the speaker is still talking.
+
+Runnable examples for both paths, hosted and local, are in
+[klang-ai/pianissimo-examples](https://github.com/klang-ai/pianissimo-examples).
+
+Pianissimo is made by [Klang](https://klang.ai/pianissimo/?utm_source=huggingface&utm_medium=referral&utm_campaign=pianissimo-subtitles&utm_content=model-card).
+
 ## Quality and speed
 
 Word error rate (WER, %) on Swedish speech; lower is better. All models were evaluated by Klang using the same scoring procedure. Bulk throughput is expressed as multiples of realtime.