freya-tr-eval-rerun / README.md
cloud0day3's picture
Alania-2 production row: best of 11 voices (12 Oct 2026, alania-v2.1)
87d38d9 verified
|
Raw History Blame Contribute Delete
7.87 kB
metadata
license: apache-2.0
language:
  - tr
task_categories:
  - text-to-speech
tags:
  - tts
  - turkish
  - benchmark
  - freya-tr-eval
pretty_name: Freya-TR-Eval re-run (PatientDesk AI)
size_categories:
  - 10K<n<100K

Freya-TR-Eval, re-run

Scripts and per-sentence results from our re-run of Freya-TR-Eval on Turkish text-to-speech systems, with one harness and the protocol published with EMA Lightning: all 495 sentences, three takes per system, an 8 kHz band, faster-whisper large-v3 at beam 5, and EMA's score_norm. Speed was measured on one NVIDIA L4 per system.

Charts, leaderboard and audio samples: https://speech.patientdesk.ai/en/benchmarks (Turkish: https://speech.patientdesk.ai/benchmarks)

Run on 7 October 2026 by PatientDesk AI; the Alania-2 production row on 12 October. Alania-2 and Alania-1 are our commercial models; Antalia-1 is our open-weights release.

Results

# System Params WER % CER % Clean UTMOS RTF (L4) Source
1 EMA Lightning (plain PyTorch) 8.6M 1.04 ± 0.17 0.22 93.7% 3.30 0.017 ours
2 EMA Lightning (compiled) 8.6M 1.04 ± 0.17 0.22 93.7% 3.30 0.0030 ours
3 Antalia-2 Mini 1.0.1 (open weights, default settings) 7.62M 1.09 ± 0.12 0.22 92.9% 3.18 0.029 ours
4 Alania-2 (production API, best of 11 voices¹) 488M 1.15 ± 0.22 (Öncel) 0.22 (Öncel) 93.1% (Lina) 3.66 (Lina) – ours
5 Trendyol-TTS (cfg 2.0, 16 steps) 2.38B 1.16 ± 0.16 0.23 93.1% 3.82 0.962 ours
6 Gemini 3.8 Flash TTS (voice Kore) – 1.23 ± 0.10 0.29 92.9% 3.53 – ours
7 Gemini 3.8 Flash-Lite TTS (voice Kore) – 1.34 ± 0.19 0.27 91.9% 3.63 – ours
8 Anka TTS (built-in male voice) 336M 1.42 ± 0.06 0.30 91.6% 3.52 0.322 ours
9 ElevenLabs v4 (voice George) – 1.43 ± 0.12 0.27 – – – quoted: EMA Lightning card
10 Alania-2 (model only, no serving checks) 488M 1.51 ± 0.19 0.30 91.8% 3.66 0.151 ours
11 ElevenLabs v4 Turbo (voice George) – 1.54 ± 0.22 0.29 – – – quoted: EMA Lightning card
12 Chatterbox Multilingual (Anka male reference) 500M 1.91 ± 0.35 0.43 88.8% 3.51 2.892 ours
13 Piper (tr_TR dfki medium, CPU) 16M 3.41 ± 0.21 0.73 79.4% 3.71 0.038 ours
14 XTTS-v2 (Anka male reference) 470M 3.67 ± 0.37 1.50 77.1% 2.97 0.455 ours
15 Alania-1 (production API) 3.3B 5.99 ± 0.47 1.81 67.3% 3.19 – ours
16 Antalia-1 (open weights, release recipe) 305M 6.28 ± 0.27 2.49 66.7% 2.72 0.476 ours
17 MMS-TTS (tur) 36M 6.38 ± 0.19 1.48 65.7% 3.79 0.014 ours
18 Coqui GlowTTS (tr common-voice, lowercased input) 28M 10.93 ± 0.00 2.96 46.1% 3.09 0.015 ours
19 FreyaTTS-small (default voice) 183M 12.02 ± 0.00 4.95 47.1% 2.81 0.286 ours
20 F5-TTS base (no Turkish training) 336M 77.84 ± 0.13 32.58 – – – quoted: Anka TTS card

¹ Alania-2 was re-run on 12 October 2026 (release alania-v2.1, production API) for all 11 built-in voices with the same harness and scoring; each column shows the best voice: Öncel for WER and CER, Lina for clean rate and UTMOS. The 11 are within noise of each other (WER 1.15–1.62%); the default voice F1 alone scored 1.38 ± 0.18 (1.44 ± 0.15 on 7 October). The other rows are from 7 October.

WER/CER: corpus level per take after score_norm, mean ± sample standard deviation over three takes. Clean: share of sentences with no word error. UTMOS: UTMOS22 strong on full-band audio (English-trained; a rough guide only). RTF: generation time / audio length on one NVIDIA L4, middle of three passes after a warm-up; API rows have none. Alania-1's RTF (2.49 on the L4, model only) was measured on the first 100 sentences. Antalia-2 Mini was re-run on 10 October 2026 at 1.0.1, which changed only its output EQ (1.0.0: 1.05 ± 0.07 / 0.20 / 93.1% / 3.25 / 0.043); its CER is 0.220 against EMA Lightning's 0.225, which both round to 0.22. Chatterbox Multilingual slowed down from pass to pass (RTF 1.69, 2.89, 4.21) with no change in output length; as for every system, the middle pass is reported. Rows marked quoted were not run by us.

What is in this repo

Path What
harness/runner.py Shared harness: warm-up, three seeded passes (sentence i of seed s uses seed 100000·s + i), WAV output, timing
harness/systems/run_*.py One adapter per open system
harness/synth_api.py Alania (our public API) and Gemini 3.8 TTS (Vertex AI), three independent requests per sentence
harness/score.py Transcription as in EMA's eval.py, plus UTMOS
harness/aggregate.py WER/CER per take, mean ± std, clean rate, UTMOS, RTF → results.json
harness/ours/ Alania-1 and Alania-2 run offline for the model-only rows; they need our private serving code and weights, and are here so the settings can be read
results/scores/<system>.jsonl Every clip: system, take, sentence id, Whisper transcript, duration, sample rate, UTMOS
results/timing/<system>.json Per-pass generation time and audio length, GPU name
results/results.json The aggregated numbers above

Audio is not included: several systems' outputs are under non-commercial licences. Everything can be regenerated with the scripts.

Reproduce

Each system gets its own virtual environment (their dependencies conflict). Set BENCH_ROOT to a working directory; the sentences are fetched from the Hub.

System Install Run
EMA Lightning 1.0.1 pip install ema-lightning soundfile run_ema.py plain / run_ema.py compiled
Antalia-2 Mini 1.0.1 pip install <wheel from huggingface.co/cloud0day3/antalia-mini> run_antalia_mini.py
Trendyol-TTS pip install voxcpm soundfile run_trendyol.py
Anka TTS v0.1 pip install "anka-tts[tts]" "pyarrow<17" soundfile run_anka.py
Chatterbox Multilingual pip install chatterbox-tts "setuptools<81" soundfile REF_WAV=ref/male.wav run_chatterbox.py
XTTS-v2, GlowTTS pip install coqui-tts "transformers<4.50" torchcodec soundfile run_xtts.py, run_glow.py
MMS-TTS pip install torch transformers soundfile run_mms.py
Piper pip install piper-tts soundfile, voice tr/tr_TR/dfki/medium from rhasspy/piper-voices PIPER_MODEL=... run_piper.py
FreyaTTS-small clone github.com/freyavoiceai/FreyaTTS, pip install -r requirements.txt FREYATTS_SRC=... run_freya.py
Antalia-1 clone github.com/0daycloud/antalia, uv sync; BigVGAN at commit 7d2b454 with the repo's patch, plus matplotlib librosa PYTHONPATH=bigvgan run_antalia.py
Scoring pip install faster-whisper librosa soundfile torch torchaudio nvidia-cublas-cu12 "nvidia-cudnn-cu12==9.*" score.py <system>... then aggregate.py

The voice reference for Chatterbox and XTTS-v2 is Anka TTS's packaged male clip (anka/data/voices/male.wav), as in the Anka TTS report. Coqui's Turkish GlowTTS has no capital letters in its symbol set and drops them silently, so run_glow.py lowercases its input the Turkish way (I→ı, İ→i); without that every sentence loses its first letter (about 21% WER).

Credits

Benchmark: Freya-TR-Eval by Freya (CC-BY-4.0). Protocol, score_norm and the transcription recipe: EMA Lightning's card and eval.py (Apache-2.0). Baseline settings for the cloning systems: the Anka TTS report. Quoted rows: ElevenLabs v4 and v4 Turbo from the EMA Lightning card; F5-TTS base from the Anka TTS card.