erayack/freya-tts-mlx-eval

Freya MLX 0.1.1: 495 paired Turkish evaluation samples

495 original Freya-TR-Eval texts, each with Torch CPU and MLX GPU synthesis: 990 original 48 kHz mono PCM16 WAVs, paired ASR transcripts and numerical-fidelity receipts. This is an evaluation-data release for runtime 0.1.1, not a new model or runtime version.

Measured result	Torch	MLX
Band-matched lexical ASR WER	444/3917 = 11.335%	442/3917 = 11.284%
CER	1019/23552 = 4.327%	1017/23552 = 4.318%

488/495 pairs have exact ASR transcript agreement. 494/495 numerical screens pass. Case0080 remains a 55.059 dB failure of the frozen ≥60 dB waveform-SNR gate. Its drift follows the latent through both decoders; expected rounding versus an upstream port defect remains unresolved. No fix or audibility conclusion is claimed.

Protocol and limits

Fixed seed 9, 32 float32 Euler steps, identical default text frontend, fallback disabled on both; Torch CPU four threads versus MLX GPU. Whisper large-v3-turbo was pinned to a4aaeec0636e6fef84abdcbe3544cb2bf7e9f6fb, with real 48→8→16 kHz bandmatching. See REPORT.md, protocol.json, ASR settings and provenance. The public protocol removes execution-budget metadata; its hash differs from the original frozen protocol, with both hashes recorded explicitly.

Semantic WER, UTMOS and human MOS were not measured. These ASR proxies do not replicate the publisher's 8% score or establish formal quality equivalence or MLX quality superiority. The historical ~65% tiny-ASR result used 20 self-written texts and a different recognizer: it is not a before/after improvement. This consumed evaluation corpus is not a future unseen test set.

Schema and use

paired_cases.jsonl has exactly 495 rows (dataset_size = 495), in original ID order. Original id, text, register, type, topic, n_words are unchanged, including nulls. source_original_text, source revision and source JSONL SHA-256 make the join explicit. torch and mlx each contain audio_path (relative WAV filename), audio_url (resolve URL), wav_sha256, pcm_sha256, transcript, spoken_target, and spoken/original word/char score counts. fidelity_screen_pass is the frozen per-case flag. Detailed float numerators and denominators are in paired-metrics.jsonl; raw ASR records are in asr.jsonl.

Ordinary JSONL and WAV are sufficient: parse each line with Python json.loads, then open row['mlx']['audio_path'] relative to this repository. The explicit URLs can also be used when constructing a datasets.Audio column (e.g. extract row['mlx']['audio_url'] into an audio column and cast it), with your own installed datasets tooling. No custom loader or parquet is required. The JSONL config is provided, but viewer/API decoding has not been validated before publication; audio URL fields are ordinary strings, not a claim of automatic Audio feature inference. Pin downloads to the returned repository commit SHA rather than mutable main; verify SHA256SUMS for every file (excluding the manifest itself).

Reconstructing results

Run python /path/to/checkout/verify.py from any working directory; no inference or packages needed. It verifies hashes, IDs, source joins, all 990 WAVs/PCM hashes, edit counts, numerical receipt arithmetic and the retained failure. REPRODUCE.md describes the evidence boundary. Full-corpus waveform/latent numerator receipts are included, not full-corpus float arrays. Float derivation cannot be recomputed from PCM16. Case0080 alone includes original paired floats and four cross-decoded NPZs; an optional NumPy check reconstructs those diagnostics.

Source, attribution and licensing

Texts are the unmodified Freya-TR-Eval, freyavoice, revision 472a2e1a19095e577fa91cbf360967934ff22b65, under CC-BY-4.0. The original source README and JSONL retain its license grant, derivation notes and full Pamuk et al. (2026), FreyaTTS Technical Report, arXiv:2607.09530 citation. Please retain that attribution when reusing the texts.

Added evaluation documentation, annotations and results are offered by erayack under CC-BY-4.0; generated audio is offered under CC-BY-4.0 to the extent of the contributor's rights. Audio is synthesized output, not original Common Voice/CoVoST recordings. No claim is made of voice-identity consent, blanket training-readiness or rights beyond the applicable grants. Review upstream provenance and applicable rights for your use; no additional upstream restriction is invented here. Model/code licenses remain separate: FreyaTTS and VoxCPM2 Apache-2.0, vendored mlx-audio MIT; these do not turn the source texts or all dataset content into Apache-2.0. Weights, runtime and source snapshots are not redistributed here. The small scoring.py and verification helpers are provided under Apache-2.0 (SPDX: Apache-2.0).