A host that keeps voices in its own store had no way to get an
embedding out of a speaker context: the encoder was only reachable
through enroll and identify. Add parakeet_capi_speaker_embed_pcm, which
reuses the helper behind enroll and returns a malloc'd L2-normalized
vector, and parakeet_capi_free_floats as its documented pair.

Non-16 kHz input is resampled as enroll does. On error the outputs are
NULL and 0 and the message is on the speaker context. Additive: keeps
ABI v10, callers probe the symbol.

Assisted-by: Claude:claude-sonnet-5-5 [Claude Code]

Check argument errors and output reset, dimension against
parakeet_capi_speaker_dim, determinism, resampling, and that one voice
scores higher than two different voices. Run it on a standalone encoder
GGUF and on a bundle voice component, and check the two give the same
embedding. The model checks skip (77) without the GGUFs.

Assisted-by: Claude:claude-sonnet-5-5 [Claude Code]

Document parakeet_capi_speaker_embed_pcm and parakeet_capi_free_floats
in speaker.md, note in bundle.md that a voice component embeds in the
same space as the standalone encoder, and add a README entry.

Assisted-by: Claude:claude-sonnet-5-5 [Claude Code]

…ording VAD evidence (#96)

* feat: add an opt-in run gate to the VAD segmenter

The Ultra and Redux heads call steady noise speech, but a noise run
sits on a plateau while a speech run sits high. Add
SegmenterOpts::run_gate (default 0 = off): a speech run, a stretch of
frames with p >= threshold found before bridging, is dropped when the
median of its frame probabilities is below the gate. A median equal to
the gate keeps the run. The gate acts first, so dropped frames are
silence for bridging, min_speech, pauses and trim. It works for any
detector and frame size, in both modes. The streaming event tracker has
no run median and ignores it.

Expose it as the "run_gate" key (a number in [0, 1)) of the VAD options
JSON, so the vad functions and the transcribe functions that take VAD
options accept it, and as --run-gate on `parakeet-cli vad` and
--vad-run-gate on `transcribe --vad`. A VAD stream refuses a non-zero
value instead of ignoring it. With the gate off the output is unchanged
byte for byte.

Tests use synthetic probability streams at 0.08 s and 0.032 s frames
with exact expected runs, the boundary, bridged gaps, a low median under
a high peak, the trim, the option parser and CLI errors. A model test
runs the gate on a clip with a seeded noise stretch for the head slice,
Ultra, Redux and Silero.

Assisted-by: Claude:claude-sonnet-5-5 [Claude Code]

* docs: the run gate, and the VAD study on real recordings

Describe the run gate in docs/vad.md: the definition of the run median,
how to use it, what it costs and what it does not do (music still
triggers a head, and it is not a noise rejector).

Add a Real recordings section to docs/vad-benchmarks.md: 59 recordings
with human labels and 2.6 h without speech. It reports frame F1 untuned
and tuned, the reference noise and a collar test, false alarms on audio
without speech, the audio the decoder receives, word error rate on clean
talks and on composites with inserted music and noise, the cost of the
30 s hard cut, and the limits.

Update the guidance. Silero stays the always-on gate, the Redux head
suits long recordings that are mostly speech, and the gate is opt-in.
The fusion gain of the synthetic experiment did not hold on real
recordings (+0.10 and +0.46 F1 points, no word error rate gain), so
fusion is not built.

Assisted-by: Claude:claude-sonnet-5-5 [Claude Code]

* scripts: add the real-recording VAD study

Add the scripts of the study (data fetch, probability dump, a Python
replica of the segmenter, frame and decoder scoring, WER reports) and the
small result tables, with a README on how to rerun them. Paths are
environment variables. verify_gate.py checks the C++ run gate against the
Python gate on stored probabilities. No audio, probabilities or models.

Assisted-by: Claude:claude-sonnet-5-5 [Claude Code]

---------

Co-authored-by: Ettore Di Giacinto <[email protected]>

fseeko and ftello are POSIX only, so the Windows cpu, cuda and vulkan
builds of v0.6.0 failed with C3861 in model_loader.cpp and the release
job was skipped. Wrap them in two small helpers that call _fseeki64 and
_ftelli64 on Windows and keep fseeko/ftello elsewhere.

Assisted-by: Claude:claude-sonnet-5-5 [Claude Code]

Co-authored-by: Ettore Di Giacinto <[email protected]>