**Bump talkies `v0.16.0` → `v0.17.0`, which adds two phoneme-recognition ASR
models. The CPU profile now serves 13 models, the CUDA profile 22.**
- Two phoneme ASR models on `/v1/audio/transcriptions`, in both the CPU and
CUDA images. They carry no language model and no lexicon, so a mispronunciation
comes back as the phones actually spoken rather than corrected to the nearest
real word. `text` is a space-separated IPA phone stream, and per-phone
timestamps come through `verbose_json`, `srt`, `vtt`, and
`timestamp_granularities` the same way the word-level models do.
- `local-talkies-wav2vec2-xlsr-53-espeak` (and its `-cuda-` alias) is a
multilingual wav2vec2 CTC model that emits eSpeak IPA, the same alphabet the
Kokoro G2P path uses.
- `local-talkies-zipa-ipa` (and its `-cuda-` alias) is a Zipformer IPA CTC
model served through the sherpa-onnx offline runtime, a 71 MB int8 download
that decodes a whole file in one pass.
- Verified end to end through LiteLLM: both return IPA for spoken input on the
CPU and CUDA routes.