You can deploy this model and start earning money today!

OpenFormosa/
BlueMagpie-TTS

Name: BlueMagpie-TTS
Rating: 2.5 (245 reviews)

Views

No views yet

BlueMagpie-TTS

BlueMagpie-TTS checkpoint for research and evaluation. The files in this Hub 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 optimizer state, scheduler state, training logs, private speaker registries, local configs, or training-data metadata.

Intended Use

Mandarin and mixed Mandarin/English text-to-speech evaluation.
Experiments with two mutually exclusive generation modes: centroid-only and speaker-reference-audio-only.
Do not redistribute the checkpoint or generated speech unless rights and consent are cleared for the intended use.

Install

bash

git clone https://github.com/voidful/BlueMagpie-TTS
cd BlueMagpie-TTS
pip install -e ".[speaker]"
pip install soundfile

Quick Start

python

from huggingface_hub import snapshot_download
import soundfile as sf

from bluemagpie import BlueMagpieModel, speaker_embedding_from_wav

model_dir = snapshot_download("OpenFormosa/BlueMagpie-TTS")
model = BlueMagpieModel.from_local(model_dir, training=False, device="cuda")

reference_embedding = speaker_embedding_from_wav(
    "reference_speaker.wav", window_sec=3.0, hop_sec=1.5
)
audio = model.generate(
    target_text="這是 AI TTS code switching 測試。",
    speaker_centroid=reference_embedding,
    cfg_value=2.0,
    inference_timesteps=10,
    retry_badcase=False,
    stop_threshold=0.65,
    stop_consecutive=2,
)

sf.write("sample.wav", audio.detach().cpu().numpy(), model.sample_rate)

Interactive Demo Endpoint Profile

The hosted BlueMagpie-TTS Demo uses the same checkpoint weights with an inference-only endpoint guard. It starts with a 0.50 stop threshold, lowers the threshold only after 75% of the native-rate duration estimate, reaches 0.05 at 95%, and accepts one stop hit. Generation is still capped at the native-rate estimate plus one latent step. Playback pace is corrected only after generation has stopped, so pace control does not hold the model open and create extra speech.

This interactive policy is intentionally separate from the reproducible offline evaluation default of stop_threshold=0.65 and stop_consecutive=2. Neither policy uses retry selection or best-of-N reranking in its standard path.

Hosted Speaker B Embedding

The hosted Space keeps Speaker A as the default and loads its fixed Speaker B voice from checkpoints/speaker_b_embedding.pt. The artifact is a 192-dimensional, L2-normalized embedding extracted from the clean Speaker B reference set. Its payload uses the bluemagpie-speaker-embedding-v1 format and includes speaker_id="female_voice", generation_seed=56789, and the Space runtime profile speaker_projector_scale=1.125.

The deployed Space pins model revision 6f7cab914a1e27c56b504ec663c0144dc25cc0a3 and validates the embedding SHA-256 e9556e14723c140985a104c1659d1ff8a5078d2fa28ce2fb756f04906641a8a7 at startup. Each text chunk is generated exactly once with retry_badcase=False; the request path does not run ASR, speaker verification, candidate reranking, or output-validation retries.

python

import os
import torch

payload = torch.load(
    os.path.join(model_dir, "checkpoints", "speaker_b_embedding.pt"),
    map_location="cpu",
    weights_only=True,
)
speaker_b = torch.nn.functional.normalize(payload["embedding"].float(), dim=0)

audio = model.generate(
    target_text="今天的會議改到下午三點。",
    speaker_centroid=speaker_b,
    cfg_value=2.0,
    inference_timesteps=10,
    retry_badcase=False,
    stop_threshold=0.65,
    stop_consecutive=2,
)

generation_seed and speaker_projector_scale are metadata for the hosted Space's fixed Speaker B runtime profile; speaker_projector_scale is not a public generate() keyword argument.

Conditioning Modes

python

# centroid-only
audio = model.generate(
    target_text="今天的會議改到下午三點。",
    speaker_centroid=centroid_tensor,
    cfg_value=2.0,
    inference_timesteps=10,
    stop_threshold=0.65,
    stop_consecutive=2,
)

# speaker-reference-audio-only: extract once; no transcript required
reference_embedding = speaker_embedding_from_wav(
    "reference_speaker.wav", window_sec=3.0, hop_sec=1.5
)
audio = model.generate(
    target_text="今天的會議改到下午三點。",
    speaker_centroid=reference_embedding,
    cfg_value=2.0,
    inference_timesteps=10,
    stop_threshold=0.65,
    stop_consecutive=2,
)

Only use reference audio or speaker centroids from speakers you have permission to synthesize. Reference clips should be at least 3 seconds. The speaker-reference-audio-only mode uses one window-averaged speaker embedding for the full request and does not require a reference transcript. Raw reference tokens remain a separately evaluated research path.

Long Text

Use the CLI for long-form synthesis. It preserves punctuation, reuses one 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 80 \
  --min-chunk-chars 12 \
  --target-chars-per-sec 4.0 \
  --stop-threshold 0.65 \
  --stop-consecutive 2 \
  --crossfade-ms 80 \
  --chunk-rms-match-db 4 \
  --chunk-edge-fade-ms 80 \
  --continuation-context-sec 0 \
  --no-retry-badcase \
  --out long.wav

Quality-First Short and Medium Text

The optional offline quality path generates multiple candidates and uses ASR plus speaker verification to select one. It is slower than ordinary generation and does not change the checkpoint weights:

bash

python scripts/generate_tts_quality.py \
  --checkpoint /path/to/model-snapshot \
  --mode speaker-reference-audio-only \
  --speaker-reference-wav /path/to/rights-cleared-reference.wav \
  --text "這是離線品質優先的語音合成測試。" \
  --candidates 10 \
  --asr-backend whisper \
  --asr-model /path/to/compatible-asr-checkpoint \
  --candidate-speaker-weight 0.05 \
  --candidate-boundary-speaker-weight 0.1 \
  --candidate-max-boundary-speaker-drop 0.03 \
  --target-chars-per-sec 4.2 \
  --out quality.wav

The candidate count and score weights are measured project heuristics, not universal TTS defaults. The command fails closed when no candidate passes the speaker-boundary gate. --allow-boundary-fallback is available for research, but is outside the measured contract. This quality policy is promoted only for speaker-reference short/medium synthesis; centroid and strict long-form gates did not pass.

Evaluation

Numbers below are from an internal held-out evaluation set. The eval set and training data are intentionally not described in this model card.

The residual-speaker pilot improved natural and long-form aggregate metrics, but every tested gate scale failed at least one stress endpoint, intelligibility, or speaker-boundary criterion. Its weights were not promoted; the selected model fingerprint remains unchanged.

A later frozen-parent endpoint-delta pilot changed only its isolated stop residual. All 24 outputs at every evaluated checkpoint remained byte-identical to the matched parent, so no endpoint-delta checkpoint was promoted. The follow-up stop-margin diagnosis is retained as research evidence and does not change the model artifact or production inference contract.

Mode	Evaluation profile	CER	WER
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.

The opt-in short/medium quality policy was measured on 30 texts with ten candidates per text. The full candidate pool reached 9.35% CER / 9.51% WER; the selected outputs reached 1.14% / 1.86%, with zero early stops, overruns, final truncations, or speaker-boundary fallbacks. Across-text speaker similarity was 0.511, reference-similarity P10 was 0.444, and median/P95 pace was 4.04/4.17 chars/sec. A second ASR that did not participate in selection measured 6.83% CER / 6.78% WER; on matched target recordings it measured 9.20% / 8.76%. The same single code-switch item was the only outlier in both. These are test-time selection results and must not be described as a trained checkpoint improvement.

The selected long-form contract was repeated in four independent processes, each with three multi-sentence texts and five request seeds. It uses 80-character chunks, 4.0 target chars/sec, no retry selection, one reference embedding per request, and no generated-audio continuation context.

Metric	Four-run range
CER	9.40%-11.90%
WER	11.13%-13.14%
Median / P95 chars per second	3.87-3.89 / 3.99
Duration CV across seeds	1.31%-2.38%
Across-text speaker similarity	0.759-0.771
Reference similarity P10	0.547-0.557
First-to-last sentence speaker drop P95	0.017-0.030
Sentence-final truncation / ASR outliers	0 / 0

A separate 12-text, one-seed stress slice reached 14.64%-15.19% CER with no final truncation; one ASCII-and-numeric-heavy item was an ASR outlier. These metrics use a corrected text-normalization contract and are not directly comparable to older long-form reports. Feeding generated audio into the next chunk consistently worsened CER, pace variance, and speaker drift, so --continuation-context-sec 0 remains the production default. Raw reference tokens are not represented by these production metrics.

A separately frozen 12-text long-form holdout reached 4.00% selected CER and 4.91% WER under the selection ASR, but one numeric/ASCII-heavy request had no candidate below the 0.03 speaker-boundary gate. Independent-ASR CER was 15.25% with one cross-chunk outlier. The quality-reranked long-form path is therefore not promoted; use the standard measured 80-character/no-context command above.

Limitations

Metrics are not a public benchmark and should be used only for internal model 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.
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.
Quality reranking is currently promoted only for speaker-reference short/medium synthesis; it is not the default centroid or long-form path.
Generated-audio continuation context is experimental and disabled by default.
Generated speech may be incorrect; do not use it as a real-world notification without human review.

Files

pytorch_model.bin: BlueMagpie model weights.
audiovae.pth: AudioVAE weights.
config.json: BlueMagpie architecture/runtime config.
tokenizer.json, tokenizer_config.json: tokenizer files.
checkpoints/*speaker_centroids.pt: small centroid demo artifacts retained on the Hub for centroid-only examples.
USAGE.md: expanded usage guide.