Link runnable hosted API recipes and clarify training partition wording

Documentation only: distinguish hosted English preview, subscription allowance and optional task pricing from local multilingual weights; link public examples. Preserve all benchmark measurements and numeric training counts.

README.md

@@ -127,7 +127,7 @@ Both models decode identical recordings with NeMo greedy-batch TDT, FP32 weights
 | Accents/domains pooled, 47 splits | 12,006 | 16.72% | **15.25%** |
 | Accents/domains English, 20 splits | 5,120 | 9.51% | **8.84%** |
 
-Orukeet improves 25 of 27 complete LibriSpeech/FLEURS splits and 36 of 47 accent/domain splits, including all 20 English accent/domain splits. The accent/domain sample contains 256 recordings per split and all 230 Lesbos recordings; the preceding FT-4035 continuation trained on 223,452 recordings across 24 complete selected partitions (371.47 hours); the 6,118 examples previously reported were a sampled follow-up fit diagnostic. Read speech and accents/domains have separate pooled results. Every recording contributes to the scores.
+Orukeet improves 25 of 27 complete LibriSpeech/FLEURS splits and 36 of 47 accent/domain splits, including all 20 English accent/domain splits. The accent/domain sample contains 256 recordings per split and all 230 Lesbos recordings; the preceding FT-4035 continuation trained on 223,452 recordings across 24 selected partitions (371.47 hours); the 6,118 examples previously reported were a sampled follow-up fit diagnostic. Read speech and accents/domains have separate pooled results. Every recording contributes to the scores.
 
 [All 74 paired WER/CER scores and edit counts](docs/current-checkpoint-benchmarks.md) · [Methods](docs/technical-report.md) · [Technical report](orukeet-technical-report.pdf)
 
@@ -225,6 +225,14 @@ Q8 and F16 pass real transcription and protocol checks on Apple silicon with Met
 
 Oruk also serves Orukeet as a hosted English transcription endpoint (preview) at [oruk.ai/models/orukeet](https://oruk.ai/models/orukeet?utm_source=huggingface&utm_medium=model-card&utm_campaign=orukeet); it needs an Oruk API key. Limits per the [OpenAPI contract](https://oruk.ai/openapi.json?utm_source=huggingface&utm_medium=model-card&utm_campaign=orukeet): English only; at most 60 seconds of audio per request; 4 MiB for the complete multipart request; no word timestamps (speaker-turn times only when diarization is enabled); usage draws from the plan's separate Orukeet allowance at $0.027 per audio hour. Hosted adds no install or GPU, plus optional emotion scores and speaker diarization on the same request. The local, open-weights path covers 25 languages and runs offline.
 
+New hosted accounts choose a subscription. Optional emotion detection and diarization have separate task rates and reduce the same Orukeet allowance; the quoted base rate does not include them. Review [plans, allowances and spending limits](https://oruk.ai/pricing#orukeet) before estimating a monthly bill.
+
+- [Python transcription](https://oruk.ai/guides/orukeet-python-transcription): install, download the sample and inspect the saved JSON output.
+- [Browser dictation](https://oruk.ai/guides/orukeet-browser-dictation): record, stop and upload through a local backend that keeps the API key private. This example does not stream partial words.
+- [Transcription with diarization](https://oruk.ai/guides/orukeet-transcription-diarization): add speaker turns with their transcript and timestamps.
+
+The [public speech API examples](https://github.com/Oruk-AI/speech-api-examples) include the runnable source and sample outputs. These are hosted-service integrations, not new local-model benchmarks.
+
 ## Questions and bugs
 
 Reproducible transcription or runtime bugs go to [GitHub issues](https://github.com/Oruk-AI/orukeet/issues/new/choose); include the model SHA-256, OS/architecture, runtime/device and input format. Model-level questions (languages, accuracy, fine-tuning, licensing, packaging Orukeet in an app) go to [GitHub Discussions](https://github.com/Oruk-AI/orukeet/discussions). Orukeet detects the language itself: the weights and the native API have no language-forcing or vocabulary-biasing input, and the native response returns `language: null`. Hosted Orukeet on oruk.ai transcribes English only. Security reports: [SECURITY.md](https://github.com/Oruk-AI/orukeet/blob/main/SECURITY.md).