| """Native conditioning for duplex inference.""" |
| import torch |
|
|
| def native_prefix(record, config): |
| """Reproduce the pinned official CustomVoice non-streaming prompt using IDs. |
| |
| A position can have projected text, codec conditioning, or their sum. |
| -1 means no embedding contribution; it is not a vocabulary token. |
| """ |
| c = config.talker_config |
| text_ids, codec_ids = [], [] |
| def add(t=-1, a=-1): |
| text_ids.append(int(t)) |
| codec_ids.append([int(a)] + [-1] * (c.num_code_groups - 1)) |
| for token in record["instruct_ids"]: |
| add(t=token) |
| ids = record["text_ids"] |
| for token in ids[:3]: |
| add(t=token) |
| lang = record["language"].lower() |
| if lang == "auto": |
| control = [c.codec_nothink_id, c.codec_think_bos_id, c.codec_think_eos_id] |
| else: |
| control = [c.codec_think_id, c.codec_think_bos_id, c.codec_language_id[lang], c.codec_think_eos_id] |
| control += [c.spk_id[record["speaker"].lower()]] |
| for token in control: |
| add(config.tts_pad_token_id, token) |
| add(config.tts_bos_token_id, c.codec_pad_id) |
| for token in ids[3:-5]: |
| add(token, c.codec_pad_id) |
| add(config.tts_eos_token_id, c.codec_pad_id) |
| add(config.tts_pad_token_id, c.codec_bos_id) |
| return torch.tensor(text_ids), torch.tensor(codec_ids) |
|
|
