Release selected stable recovery checkpoint

README.md

@@ -14,9 +14,8 @@ license: other
 # BlueMagpie-TTS
 
 BlueMagpie-TTS checkpoint for research and evaluation. The files in this Hub
-repository are the selected balanced two-mode release. Newer experimental
-checkpoints are not promoted unless they pass the same intelligibility,
-duration, endpoint, and speaker-stability gates.
+repository are the selected two-mode recovery release. Experimental branches
+are not promoted when they regress the production speaker-reference path.
 
 This repository contains the inference artifact only: model weights, AudioVAE
 weights, tokenizer files, config, and usage documentation. It does not include
@@ -60,6 +59,7 @@ audio = model.generate(
     cfg_value=2.0,
     inference_timesteps=10,
     retry_badcase=True,
+    retry_badcase_ratio_threshold=6.0,
     stop_threshold=0.65,
     stop_consecutive=2,
 )
@@ -108,7 +108,7 @@ speaker embedding for every chunk, applies punctuation-aware pauses, and can
 set a minimum duration from the target speaking rate:
 
 ```bash
-python scripts/generate_tts.py   --checkpoint /path/to/model-snapshot   --mode speaker-reference-audio-only   --speaker-reference-wav /path/to/rights-cleared-reference.wav   --reference-audio-conditioning speaker-embedding   --text-file long_text.txt   --chunk-chars 45   --min-chunk-chars 12   --target-chars-per-sec 4.2   --stop-threshold 0.65   --stop-consecutive 2   --crossfade-ms 80   --out long.wav
+python scripts/generate_tts.py   --checkpoint /path/to/model-snapshot   --mode speaker-reference-audio-only   --speaker-reference-wav /path/to/rights-cleared-reference.wav   --reference-audio-conditioning speaker-embedding   --text-file long_text.txt   --chunk-chars 120   --min-chunk-chars 12   --target-chars-per-sec 4.0   --stop-threshold 0.65   --stop-consecutive 2   --crossfade-ms 80   --out long.wav
 ```
 
 ## Evaluation
@@ -116,26 +116,35 @@ python scripts/generate_tts.py   --checkpoint /path/to/model-snapshot   --mode s
 Numbers below are from an internal held-out evaluation set. The eval set and
 training data are intentionally not described in this model card.
 
-| System | Setting | CER | WER |
+| Mode | Evaluation profile | CER | WER |
 | --- | --- | ---: | ---: |
-| BlueMagpie-TTS | centroid-only | 7.44% | 8.57% |
-| BlueMagpie-TTS | raw-reference-tokens research path | 8.99% | 11.77% |
-| BlueMagpie-TTS | speaker-reference-embedding production path | pending re-qualification | pending re-qualification |
-| Reference baseline | same internal eval | 11.45% | 14.83% |
-
-Long-form diagnostics on the selected checkpoint:
-
-| Metric | Value |
-| --- | ---: |
-| centroid-only CER / WER | 7.51% / 7.51% |
-| raw-reference-tokens CER / WER | 8.92% / 8.92% |
-| Median chars/sec | 4.017 |
-| speaker-reference chunk speaker-sim drop | 0.109 |
-
-The measured reference-audio rows above used the raw-reference-token research
-backend. They are not presented as production speaker-embedding results. The
-long-form rate was measured with controlled chunked inference; it does not show
-that unrestricted single-pass endpoint behavior is fully solved.
+| centroid-only | 30 texts x 5 seeds, controlled pace | 15.37% | 41.48% |
+| speaker-reference-embedding | 30 texts x 5 seeds, controlled pace | 13.39% | 33.05% |
+
+The controlled profile uses 4.2 target chars/sec, stop threshold 0.65, two-hit
+hysteresis, and no retry selection. Its median/P95 pace is 4.04/4.17 chars/sec.
+The stricter natural-stop profile has lower CER but fails the pace gate: P95
+reaches 6.25 chars/sec and early-stop classification exceeds 50%. Controlled
+pace is therefore part of the current production contract, not an optional
+benchmark optimization.
+
+Long-form diagnostics use 12 varied texts and five seeds. The final reference
+path calibration uses 4.0 target chars/sec; the centroid comparison retains
+the matched 4.2 target used during checkpoint selection:
+
+| Mode | Path | Target CPS | CER | WER | Across-text speaker similarity |
+| --- | --- | ---: | ---: | ---: | ---: |
+| centroid-only | direct | 4.2 | 24.51% | 55.79% | 0.619 |
+| centroid-only | chunked, 120 chars | 4.2 | 27.16% | 61.83% | 0.628 |
+| speaker-reference-embedding | direct | 4.0 | 23.33% | 52.64% | 0.608 |
+| speaker-reference-embedding | chunked, 120 chars | 4.0 | 26.54% | 56.13% | 0.625 |
+
+Chunking improves speaker consistency and sentence-final completeness but can
+trade reference-mode CER. At 4.0 target chars/sec, reference chunking reduced
+the first-to-last sentence speaker drop from 0.108 at the previous 4.2 policy
+to 0.025 and reduced final truncation from 18.33% to 16.67%. Use it for
+multi-sentence or long input. Raw reference tokens are not represented by
+these production metrics.
 
 ## Limitations
 
@@ -143,11 +152,10 @@ that unrestricted single-pass endpoint behavior is fully solved.
   selection.
 - Speaker similarity depends on the quality and rights-cleared status of the
   supplied reference audio or centroid.
-- The windowed speaker-embedding backend is the recommended production path,
-  but its full release-gate metrics are still pending re-qualification.
-- Free-running endpoint behavior can still produce rare early stops, overruns,
-  repetitions, or extra speech. Keep retry and stop hysteresis enabled for
-  offline generation and inspect important outputs.
+- The windowed speaker-embedding backend is the recommended production path.
+- Free-running endpoint behavior is not fully calibrated. Use the documented
+  target pace, stop hysteresis, and chunking contract; inspect important output.
+- Raw reference tokens remain a separately gated research backend.
 - Very long passages should be chunked to avoid stop-token and prosody drift.
 - Generated speech may be incorrect; do not use it as a real-world notification
   without human review.

USAGE.md

@@ -33,6 +33,7 @@ wav = model.generate(
     cfg_value=2.0,
     inference_timesteps=10,
     retry_badcase=True,
+    retry_badcase_ratio_threshold=6.0,
     stop_threshold=0.65,
     stop_consecutive=2,
 )
@@ -57,7 +58,7 @@ reference_embedding = speaker_embedding_from_wav(
     "reference_speaker.wav", window_sec=3.0, hop_sec=1.5
 )
 wav = model.generate(
-    target_text="明天 meeting 取消一次，我們大家放假。",
+    target_text="今天的 meeting 依照原定時間進行。",
     speaker_centroid=reference_embedding,
     cfg_value=2.0,
     inference_timesteps=10,
@@ -103,7 +104,7 @@ sf.write("streaming.wav", wav.numpy(), model.sample_rate)
 Long text:
 
 ```bash
-python scripts/generate_tts.py   --checkpoint /path/to/model-snapshot   --mode speaker-reference-audio-only   --speaker-reference-wav /path/to/rights-cleared-reference.wav   --reference-audio-conditioning speaker-embedding   --text-file long_text.txt   --chunk-chars 45   --min-chunk-chars 12   --target-chars-per-sec 4.2   --stop-threshold 0.65   --stop-consecutive 2   --crossfade-ms 80   --out long.wav
+python scripts/generate_tts.py   --checkpoint /path/to/model-snapshot   --mode speaker-reference-audio-only   --speaker-reference-wav /path/to/rights-cleared-reference.wav   --reference-audio-conditioning speaker-embedding   --text-file long_text.txt   --chunk-chars 120   --min-chunk-chars 12   --target-chars-per-sec 4.0   --stop-threshold 0.65   --stop-consecutive 2   --crossfade-ms 80   --out long.wav
 ```
 
 This path extracts the windowed speaker embedding once and reuses it for every
@@ -115,9 +116,10 @@ Recommended generation defaults:
 - `cfg_value=2.0`
 - `inference_timesteps=10`
 - `retry_badcase=True` for offline generation
+- `retry_badcase_ratio_threshold=6.0` so retry detection is not a pace cap
 - `stop_threshold=0.65`, `stop_consecutive=2` for offline generation
 - `max_len=2000` unless a longer chunk is intentionally needed
-- `target-chars-per-sec=4.2` for controlled long-form chunking
+- `target-chars-per-sec=4.0` for controlled long-form chunking
 - Reference wavs used to extract speaker embeddings should be at least 3 seconds.
 
 Safety:

config.json

@@ -242,11 +242,22 @@
         0.25,
         0.75
       ],
+      "time_logit_mean": -0.4,
+      "time_logit_std": 1.0,
       "noise_cond_prob_range": [
         0.0,
         0.0
       ],
-      "noise_cond_scale": 0.0
+      "noise_cond_scale": 0.0,
+      "full_interval_prob": 0.0,
+      "full_interval_mode": "absolute",
+      "full_interval_r": 0.0,
+      "full_interval_t": 1.0,
+      "integrated_cfg_training": false,
+      "integrated_cfg_inference": false,
+      "integrated_cfg_omega": 0.3,
+      "integrated_cfg_kappa": 0.9,
+      "cfg_rescale_phi": 0.0
     }
   },
   "audio_vae_config": {
@@ -287,6 +298,9 @@
   },
   "speaker_embed_dim": 192,
   "ref_audio_bottleneck": "none",
+  "ref_audio_summary_tokens": 4,
+  "speaker_step_injection": true,
+  "speaker_step_injection_init": "zero",
   "audio_start_token": -1,
   "audio_end_token": -1,
   "ref_audio_start_token": -1,

pytorch_model.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:e454c077ded2ce9d5f3b8ec4f336af840a2fe9912149d1f25afd4938bf7a0319
-size 7746547131
+oid sha256:4c2f67a50db73a8bd4e35f6ecef7e9c874c13e22388a0e9dc73330e4a679d7dd
+size 7783283863

release_metadata.json

@@ -1,6 +1,6 @@
 {
   "format": "bluemagpie-pretrained",
-  "checkpoint": "step_0006000",
+  "checkpoint": "step_0000100",
   "release_status": "selected_stable",
   "files": [
     "pytorch_model.bin",
@@ -11,7 +11,7 @@
   ],
   "artifacts": {
     "pytorch_model.bin": {
-      "sha256": "e454c077ded2ce9d5f3b8ec4f336af840a2fe9912149d1f25afd4938bf7a0319"
+      "sha256": "4c2f67a50db73a8bd4e35f6ecef7e9c874c13e22388a0e9dc73330e4a679d7dd"
     },
     "audiovae.pth": {
       "sha256": "cc01a2d5d817910f185d959f8bb23d1a5ab80c9a03b7398ac86d7457cb663332"
@@ -29,35 +29,42 @@
   "evaluation": {
     "scope": "internal held-out evaluation",
     "centroid_only": {
-      "cer": 0.07441389020336389,
-      "wer": 0.08565566378066376
-    },
-    "speaker_reference_audio_only": {
-      "backend": "raw_reference_tokens",
-      "release_role": "research",
-      "cer": 0.08992608978902839,
-      "wer": 0.11774568024568023
+      "profile": "controlled_30_texts_5_seeds",
+      "cer": 0.1537326315659649,
+      "wer": 0.4148126965626966,
+      "chars_per_second_median": 4.044117647058823,
+      "chars_per_second_p95": 4.166666666666667,
+      "speaker_across_texts_median": 0.4630973200859695,
+      "speaker_centroid_p10": 0.37282803654670715
     },
     "speaker_reference_embedding": {
       "release_role": "production",
-      "status": "pending_requalification"
-    },
-    "long_centroid_only": {
-      "cer": 0.07511737089201878,
-      "wer": 0.07511737089201878,
-      "chars_per_second": 4.017351942663145
+      "profile": "controlled_30_texts_5_seeds",
+      "cer": 0.13387449587449585,
+      "wer": 0.330523828023828,
+      "chars_per_second_median": 4.044117647058823,
+      "chars_per_second_p95": 4.166666666666667,
+      "duration_cv": 0.045454545454545414,
+      "speaker_across_texts_median": 0.4557993102947186,
+      "speaker_reference_p10": 0.3619180113077164
     },
-    "long_speaker_reference_audio_only": {
-      "backend": "raw_reference_tokens",
-      "release_role": "research",
-      "cer": 0.0892018779342723,
-      "wer": 0.0892018779342723,
-      "chars_per_second": 4.017351942663145,
-      "chunk_speaker_similarity_drop": 0.1085478663444519
+    "natural_stop": {
+      "status": "pace_gate_failed",
+      "speaker_reference_cer": 0.1285453712953713,
+      "speaker_reference_p95_chars_per_second": 6.25,
+      "speaker_reference_early_stop_rate": 0.5066666666666667
     },
-    "reference_baseline": {
-      "cer": 0.114483,
-      "wer": 0.148266
+    "longform_chunked_120": {
+      "profile": "12_texts_5_seeds",
+      "centroid_target_chars_per_second": 4.2,
+      "centroid_cer": 0.27156543786258974,
+      "centroid_speaker_across_texts_median": 0.6284434280612252,
+      "speaker_reference_target_chars_per_second": 4.0,
+      "speaker_reference_cer": 0.26542808108820687,
+      "speaker_reference_wer": 0.5612998112998112,
+      "speaker_reference_across_texts_median": 0.6251172526316209,
+      "speaker_reference_final_truncation_rate": 0.16666666666666666,
+      "speaker_reference_first_last_sentence_drop_p95": 0.024524328112602203
     }
   }
 }