v0.3.0

Highlights

• Speech-LLM transcription with Nemotron 3.5 Transcribe 30B-A4B. Transcribes complete recordings of up to one hour with punctuation and capitalization, and adds word timestamps, speaker labels, speech translation and supports custom instructions. It runs through nemo-speech transcribe, the HTTP server (including /v1/chat/completions) and riva_server. The default CUDA build decodes many requests together, with MTP or DFlash2 speculative decoding. See the SALM guide. (#82, #84, #85)
• Speaker labels in text, SRT and WebVTT. With --diarize, text output has one Speaker N: ... paragraph per turn, SRT cues start with Speaker N: and WebVTT cues use <v Speaker N>. A cue never spans a speaker change, and the CLI and HTTP server share one subtitle renderer. (#87)
• Zero-copy ASR weights. ASR models memory-map their GGUF and bind weights directly on CPU and Metal, instead of holding a second copy in memory. (#47)
• MagpieTTS and NanoCodec on integrated NVIDIA GPUs. On unified-memory devices such as DGX Spark, TTS now runs on the GPU instead of silently falling back to the CPU. (#65)
• Faster offline diarization front end. The mel extractor computes the power spectrum once per frame instead of once per mel band, and stream buffers are compacted lazily. (#62)

Fixes

• Streaming ASR no longer drops speech at the start of a stream, and Q8 pointwise layers keep their precision. (#46)
• Word times stay inside the audio: the end-of-stream flush could place them up to 1.3 s past the end. Words at speaker changes on cache-aware models now get the right speaker. (#87)
• Long recordings without endpointing no longer collapse to one speaker after about 10 minutes (word speaker accuracy on AMI ES2002a: 59.2% → 99.7%). (#70)
• MagpieTTS v2607 conversion maps speakers to the right ids. (#66)
• The NanoCodec decoder fails a decode that produces NaN or infinite samples, instead of returning them clamped to full-scale audio. (#43)
• Text normalization no longer inserts "sil" for an isolated comma; OpenFST and Sparrowhawk come from updated repositories. (#69)

Notes

• cuBLAS in the Linux CUDA archives and the container. They now ship NVIDIA's cuBLAS and cuBLASLt instead of the in-tree shim due to performance concerns, so they are larger. (#88)

New Contributors

• @dashidhy made their first contribution in #46
• @Eoin-Houstoun made their first contribution in #62
• @femo23 made their first contribution in #65

Full Changelog: v0.2.0...v0.3.0