enne2/qwentts.cpp
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

cli, facade: rename --ref-audio to --ref-wav, align with omnivoice convention

Browse Source
This commit is contained in:
Pascal
2026-05-14 15:46:46 +02:00
parent 6ad6a7db7a
commit b1e864cb8d
7 changed files with 26 additions and 27 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/pipeline-tts.cpp
+4 -5
View File
@@ -288,8 +288,7 @@ bool pipeline_tts_synthesize(PipelineTTS * pt,
const float * ref_spk_emb_ptr = NULL;
if (has_ref_audio) {
if (!pt->has_speaker_encoder) {
fprintf(stderr,
"[Pipeline] FATAL: --ref-audio requires a model with a loaded speaker encoder (Base only)\n");
fprintf(stderr, "[Pipeline] FATAL: --ref-wav requires a model with a loaded speaker encoder (Base only)\n");
return false;
}
if (!speaker_encoder_extract(&pt->speaker_encoder, pt->sched, params.ref_audio_24k, params.ref_n_samples,
@@ -313,13 +312,13 @@ bool pipeline_tts_synthesize(PipelineTTS * pt,
int ref_codes_T = 0;
if (!ref_text.empty()) {
if (!has_ref_audio) {
fprintf(stderr, "[Pipeline] FATAL: --ref-text requires --ref-audio\n");
fprintf(stderr, "[Pipeline] FATAL: --ref-text requires --ref-wav\n");
return false;
}
// The codec hop is 1920 samples at 24 kHz so n_samples must be
// a multiple of 1920. Truncate to the nearest hop boundary.
if (params.ref_n_samples < QWEN_TOKENIZER_HOP_LENGTH) {
fprintf(stderr, "[Pipeline] FATAL: ref_audio too short for ICL (%d samples)\n", params.ref_n_samples);
fprintf(stderr, "[Pipeline] FATAL: ref_wav too short for ICL (%d samples)\n", params.ref_n_samples);
return false;
}
int aligned_T = (params.ref_n_samples / QWEN_TOKENIZER_HOP_LENGTH) * QWEN_TOKENIZER_HOP_LENGTH;
@@ -347,7 +346,7 @@ bool pipeline_tts_synthesize(PipelineTTS * pt,
debug_dump_2d(&d, "trailing-text-hidden", prompt.trailing_text_hidden.data(), prompt.T_trailing, prompt.hidden);
debug_dump_1d(&d, "tts-pad-embed", prompt.tts_pad_embed.data(), prompt.hidden);
// Voice clone dumps : speaker-emb fires when ref_audio is set
// Voice clone dumps : speaker-emb fires when ref_wav is set
// (modes A and B), ref-codes fires only when ref_text is also set
// (mode B ICL). Both are no-ops in base / tts / customvoice modes,
// the dump files simply do not appear in those runs.
src/prompt-builder.cpp
+1 -1
View File
@@ -215,7 +215,7 @@ bool prompt_builder_build(const PipelineTTS * pt,
// Mode B requires ref_spk_emb so the speaker slot is also filled.
const bool icl = !ref_text.empty() && ref_codes != NULL && ref_codes_T > 0;
if (icl && ref_spk_emb == NULL) {
fprintf(stderr, "[Prompt] FATAL: ICL mode requires ref_spk_emb (no --ref-audio?)\n");
fprintf(stderr, "[Prompt] FATAL: ICL mode requires ref_spk_emb (no --ref-wav?)\n");
return false;
}
src/qwen.cpp
+3 -3
View File
@@ -328,17 +328,17 @@ enum qwen_status qwen_synthesize(struct qwen_context * q,
return QWEN_STATUS_MODE_INVALID;
}
if (params->ref_audio_24k && mt != "base") {
qt_set_error("--ref-audio is only valid for base models (loaded: %s)", mt.c_str());
qt_set_error("--ref-wav is only valid for base models (loaded: %s)", mt.c_str());
qwen_audio_free(out);
return QWEN_STATUS_MODE_INVALID;
}
if (params->speaker && params->ref_audio_24k) {
qt_set_error("--speaker and --ref-audio are mutually exclusive");
qt_set_error("--speaker and --ref-wav are mutually exclusive");
qwen_audio_free(out);
return QWEN_STATUS_INVALID_PARAMS;
}
if (params->ref_text && !params->ref_audio_24k) {
qt_set_error("--ref-text requires --ref-audio");
qt_set_error("--ref-text requires --ref-wav");
qwen_audio_free(out);
return QWEN_STATUS_INVALID_PARAMS;
}