Skip to content

[TTS] Allow inference without reference audio #15213

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
rfejgin merged 7 commits into from
Jan 7, 2026
+68 −39
Merged

[TTS] Allow inference without reference audio #15213

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
7 commits
Select commit Hold shift + click to select a range
a5ac86a
Added datasets filtering to the inference script
rfejgin Dec 19, 2025
10a838b
Refined datasets filtering in the inference script
rfejgin Dec 20, 2025
d2fa1ea
Allow inference without reference audio
rfejgin Dec 20, 2025
7c4098b
Remove changes not related to this PR
rfejgin Dec 22, 2025
683a750
Merge remote-tracking branch 'nemo/main' into magpietts_inference_wit…
rfejgin Jan 5, 2026
9301912
Address PR comments
rfejgin Jan 6, 2026
1827143
Correct an assertion per PR comments
rfejgin Jan 6, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
15 changes: 8 additions & 7 deletions nemo/collections/tts/data/text_to_speech_dataset.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -459,7 +459,7 @@ def __getitem__(self, index):
if 'audio_filepath' in data.manifest_entry:
# If audio_filepath is available, then use the actual audio file path.
example['audio_filepath'] = data.manifest_entry['audio_filepath']
else:
elif 'audio_filepath' in data.manifest_entry:
# Only load audio if codes are not available
audio_array, _, audio_filepath_rel = load_audio(
manifest_entry=data.manifest_entry,
Expand Down Expand Up @@ -661,13 +661,15 @@ def collate_fn(self, batch: List[dict]):
speaker_indices_list = []
for example in batch:
dataset_name_list.append(example["dataset_name"])
audio_filepath_list.append(example["audio_filepath"])
raw_text_list.append(example["raw_text"])
language_list.append(example["language"])

token_list.append(example["tokens"])
token_len_list.append(example["text_len"])

if 'audio_filepath' in example:
audio_filepath_list.append(example["audio_filepath"])

if 'audio' in example:
audio_list.append(example["audio"])
audio_len_list.append(example["audio_len"])
Expand Down Expand Up @@ -774,14 +776,13 @@ def collate_fn(self, batch: List[dict]):
if len(speaker_indices_list) > 0:
batch_dict['speaker_indices'] = torch.tensor(speaker_indices_list, dtype=torch.int64)

# Assert only ONE of context_audio or context_audio_codes in the batch
assert ('audio' in batch_dict) ^ ('audio_codes' in batch_dict)
# Assert no more than one of audio or audio_codes in the batch
if 'audio' in batch_dict:
assert 'audio_codes' not in batch_dict

# Assert only ONE of context_audio or context_audio_codes in the batch
# Assert no more than one of context_audio or context_audio_codes in the batch
if 'context_audio' in batch_dict:
assert 'context_audio_codes' not in batch_dict
if 'context_audio_codes' in batch_dict:
assert 'context_audio' not in batch_dict

return batch_dict

Expand Down
92 changes: 60 additions & 32 deletions nemo/collections/tts/modules/magpietts_inference/evaluate_generated_audio.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -246,9 +246,9 @@ def evaluate(
gt_audio_texts = []
total_generated_audio_seconds = 0.0
for ridx, record in enumerate(records):
gt_audio_filepath = record['audio_filepath']
gt_audio_filepath = record.get('audio_filepath', None)
context_audio_filepath = record.get('context_audio_filepath', None)
if audio_dir is not None:
if audio_dir is not None and gt_audio_filepath is not None:
gt_audio_filepath = os.path.join(audio_dir, gt_audio_filepath)
if context_audio_filepath is not None:
context_audio_filepath = os.path.join(audio_dir, context_audio_filepath)
Expand All @@ -265,17 +265,25 @@ def evaluate(
with torch.inference_mode():
pred_text = asr_model.transcribe([pred_audio_filepath], batch_size=1, use_lhotse=False)[0].text
pred_text = process_text(pred_text)
gt_audio_text = asr_model.transcribe([gt_audio_filepath], batch_size=1, use_lhotse=False)[0].text
gt_audio_text = process_text(gt_audio_text)
if gt_audio_filepath is not None:
gt_audio_text = asr_model.transcribe([gt_audio_filepath], batch_size=1, use_lhotse=False)[
0
].text
gt_audio_text = process_text(gt_audio_text)
else:
gt_audio_text = None
else:
pred_text = transcribe_with_whisper(
whisper_model, whisper_processor, pred_audio_filepath, language, device
)
pred_text = process_text(pred_text)
gt_audio_text = transcribe_with_whisper(
whisper_model, whisper_processor, gt_audio_filepath, language, device
)
gt_audio_text = process_text(gt_audio_text)
if gt_audio_filepath is not None:
gt_audio_text = transcribe_with_whisper(
whisper_model, whisper_processor, gt_audio_filepath, language, device
)
gt_audio_text = process_text(gt_audio_text)
else:
gt_audio_text = None
except Exception as e:
logging.info("Error during ASR: {}".format(e))
pred_text = ""
Expand Down Expand Up @@ -318,19 +326,29 @@ def evaluate(
sv_model_type=sv_model_type,
)

# Ground truth vs. predicted
gt_speaker_embedding = extract_embedding_fn(audio_path=gt_audio_filepath)
pred_speaker_embedding = extract_embedding_fn(audio_path=pred_audio_filepath)
pred_gt_ssim = torch.nn.functional.cosine_similarity(
gt_speaker_embedding, pred_speaker_embedding, dim=0
).item()
# Initialize SSIMs with a default since the context or ground truth audio
# may be unavailable.
pred_context_ssim = float('NaN')
gt_context_ssim = float('NaN')
pred_context_ssim_alternate = float('NaN')
gt_context_ssim_alternate = float('NaN')
pred_gt_ssim = float('NaN')
pred_gt_ssim_alternate = float('NaN')

if gt_audio_filepath is not None:
# Ground truth vs. predicted
gt_speaker_embedding = extract_embedding_fn(audio_path=gt_audio_filepath)
pred_speaker_embedding = extract_embedding_fn(audio_path=pred_audio_filepath)
pred_gt_ssim = torch.nn.functional.cosine_similarity(
gt_speaker_embedding, pred_speaker_embedding, dim=0
).item()

# Ground truth vs. predicted (alternate model)
gt_speaker_embedding_alternate = extract_embedding_fn_alternate(audio_path=gt_audio_filepath)
pred_speaker_embedding_alternate = extract_embedding_fn_alternate(audio_path=pred_audio_filepath)
pred_gt_ssim_alternate = torch.nn.functional.cosine_similarity(
gt_speaker_embedding_alternate, pred_speaker_embedding_alternate, dim=0
).item()
# Ground truth vs. predicted (alternate model)
gt_speaker_embedding_alternate = extract_embedding_fn_alternate(audio_path=gt_audio_filepath)
pred_speaker_embedding_alternate = extract_embedding_fn_alternate(audio_path=pred_audio_filepath)
pred_gt_ssim_alternate = torch.nn.functional.cosine_similarity(
gt_speaker_embedding_alternate, pred_speaker_embedding_alternate, dim=0
).item()

if context_audio_filepath is not None:
context_speaker_embedding = extract_embedding_fn(audio_path=context_audio_filepath)
Expand All @@ -341,18 +359,20 @@ def evaluate(
pred_speaker_embedding, context_speaker_embedding, dim=0
).item()
# Ground truth vs. context
gt_context_ssim = torch.nn.functional.cosine_similarity(
gt_speaker_embedding, context_speaker_embedding, dim=0
).item()
if gt_audio_filepath is not None:
gt_context_ssim = torch.nn.functional.cosine_similarity(
gt_speaker_embedding, context_speaker_embedding, dim=0
).item()

# Predicted vs. context (alternate model)
pred_context_ssim_alternate = torch.nn.functional.cosine_similarity(
pred_speaker_embedding_alternate, context_speaker_embedding_alternate, dim=0
).item()
# Ground truth vs. context (alternate model)
gt_context_ssim_alternate = torch.nn.functional.cosine_similarity(
gt_speaker_embedding_alternate, context_speaker_embedding_alternate, dim=0
).item()
if gt_audio_filepath is not None:
gt_context_ssim_alternate = torch.nn.functional.cosine_similarity(
gt_speaker_embedding_alternate, context_speaker_embedding_alternate, dim=0
).item()
total_generated_audio_seconds += get_wav_file_duration(pred_audio_filepath)

filewise_metrics.append(
Expand Down Expand Up @@ -415,12 +435,20 @@ def evaluate(
avg_metrics['ssim_gt_context_avg_alternate'] = sum(
[m['gt_context_ssim_alternate'] for m in filewise_metrics]
) / len(filewise_metrics)
avg_metrics["cer_gt_audio_cumulative"] = word_error_rate_detail(
hypotheses=gt_audio_texts, references=gt_texts, use_cer=True
)[0]
avg_metrics["wer_gt_audio_cumulative"] = word_error_rate_detail(
hypotheses=gt_audio_texts, references=gt_texts, use_cer=False
)[0]
if not None in gt_audio_texts:
avg_metrics["cer_gt_audio_cumulative"] = word_error_rate_detail(
hypotheses=gt_audio_texts, references=gt_texts, use_cer=True
)[0]
avg_metrics["wer_gt_audio_cumulative"] = word_error_rate_detail(
hypotheses=gt_audio_texts, references=gt_texts, use_cer=False
)[0]
else:
avg_metrics["cer_gt_audio_cumulative"] = float('NaN')
avg_metrics["wer_gt_audio_cumulative"] = float('NaN')
logging.warning(
"Ground truth audio files are missing. Setting cumulative CER and WER for ground truth audio to NaN."
)

avg_metrics["utmosv2_avg"] = sum([m['utmosv2'] for m in filewise_metrics]) / len(filewise_metrics)
avg_metrics["total_gen_audio_seconds"] = total_generated_audio_seconds
pprint.pprint(avg_metrics)
Expand Down
Loading