---
license: cc-by-4.0
language:
- tr
task_categories:
- automatic-speech-recognition
tags:
- turkish
- speech-recognition
- benchmark
- wer
- cer
- real-time-factor
pretty_name: TurkMedSTT Turkish ASR Benchmark
size_categories:
- 10K<n<100K
---

# TurkMedSTT Turkish ASR Benchmark

This repository publishes general-domain Turkish ASR benchmark metrics produced by
Muhammed Kumcu and Yagmur Tuncer.

**Interactive table:**
[`turkmedstt/turkish-asr-leaderboard`](https://huggingface.co/spaces/turkmedstt/turkish-asr-leaderboard)

The release compares 22 ASR models on general-domain Turkish speech clips drawn from
three established Turkish speech sources. It contains metrics only. Source audio,
reference transcripts, generated hypotheses, local paths, and medical evaluation
results are not distributed.

## Evaluation protocol

- Each row reports its number of evaluated general-domain clips in the Clips column.
- Each model produces one hypothesis for each clip.
- Text normalization and metric computation follow the same benchmark pipeline.
- Overall values are macro averages over clips, so every clip contributes equally.
- The primary ranking key is mean WER. CER and RTF are supporting metrics.
- AcoSemantic values are complementary analysis, not a replacement for accuracy.
- RTF comes from the original Colab A100 runs and depends on backend, model
  implementation, batching, hardware, and decoding choices.

## Dataset composition

| Source | Clips | Role in the benchmark |
|---|---:|---|
| Common Voice Turkish | 447 | Multi-speaker community recordings and short general utterances |
| ISSAI Turkish | 453 | Read-speech recordings with different sentence and speaker characteristics |
| OpenSLR 106 Turkish | 160 | Short Turkish utterances from a separately distributed speech corpus |
| **Total** | **1,060** | About 105 minutes of general-domain Turkish speech |

The release does not redistribute source audio or transcripts. Dataset names and
de-identified clip identifiers are retained only to describe provenance and join
metric tables. Original source licenses and access conditions continue to apply.

## Column guide

| Column | Meaning | Direction and example |
|---|---|---|
| Rank | Position after sorting by mean WER | Lower rank number is better in the primary ranking |
| Model | Hugging Face model identifier | Identifies the evaluated checkpoint |
| Backend | Inference implementation used in the run | Backends can affect speed and decoding behavior |
| Clips | Number of evaluated general-domain clips | Per-row count; most rows use 1,060 clips |
| WER | Word error rate: `(substitutions + deletions + insertions) / reference words` | Lower is better. `0.1345` means about 13.45 word errors per 100 reference words |
| CER | Character error rate using the same edit operations at character level | Lower is better; useful for Turkish suffix and spelling differences |
| RTF | Real-time factor: `processing seconds / audio seconds` | Lower is faster. `0.10` means roughly 10x faster than real time |
| ASCS | `1 - abs(reference sentiment - hypothesis sentiment)` | Higher is better; measures preservation of text-level semantic-affective score |
| ASCS sigma | Population standard deviation of clip-level ASCS values | Lower means more consistent ASCS behavior across clips |
| ASCS-WER r | Pearson correlation between clip-level WER and ASCS | Negative is expected: more transcription error generally accompanies lower ASCS |
| Reference sentiment | Mean sentiment score of reference transcripts | Describes the test distribution; it is not a model quality score |
| Hypothesis sentiment | Mean sentiment score of model outputs | Compare with reference sentiment to inspect systematic affective shift |
| Sentiment drift | Mean absolute difference between reference and hypothesis sentiment | Lower is better; numerically equal to `1 - mean ASCS` in this definition |

WER and CER may exceed `1.0` on individual clips when a model inserts many extra
tokens. A low RTF does not imply high transcription quality, and a high ASCS does
not imply that the words are correct. The metrics must be interpreted together.

## Full leaderboard

WER and CER are macro means over clips. RTF is the mean real-time factor measured
in the original Colab A100 runs. ASCS_text measures text-level
semantic-affective consistency; higher is better. ASCS-WER `r` is the per-model
Pearson correlation across clips. A negative value means higher WER generally
coincides with lower semantic-affective consistency.

| Rank | Model | Backend | WER ↓ | CER ↓ | RTF ↓ | ASCS ↑ | ASCS σ ↓ | ASCS-WER r | Sentiment drift ↓ |
|---:|---|---|---:|---:|---:|---:|---:|---:|---:|
| 1 | `turkmedstt/whisper-large-v3-turkish-general` | transformers | 0.0777 | 0.0234 | 0.1465 | 0.9832 | 0.0328 | -0.3836 | 0.0168 |
| 2 | `turkmedstt/whisper-large-v3-turkish-medical` | transformers | 0.0782 | 0.0233 | 0.1488 | 0.9833 | 0.0328 | -0.3854 | 0.0167 |
| 3 | `openai/whisper-large-v3` | transformers | 0.1345 | 0.0588 | 0.1349 | 0.9571 | 0.0563 | -0.3471 | 0.0429 |
| 4 | `vincespeed/faster-whisper-large-v3-turbo-turkish` | faster-whisper | 0.1825 | 0.0774 | 0.0927 | 0.9534 | 0.0579 | -0.1874 | 0.0466 |
| 5 | `Huseyin/whisper-large-v3-turkish-finetuned` | transformers | 0.1841 | 0.0971 | 0.1587 | 0.9510 | 0.0772 | -0.4160 | 0.0472 |
| 6 | `openai/whisper-large-v2` | transformers | 0.1901 | 0.0999 | 0.1333 | 0.9498 | 0.0633 | -0.3140 | 0.0502 |
| 7 | `openai/whisper-large-v3-turbo` | transformers | 0.2014 | 0.0897 | 0.0338 | 0.9487 | 0.0687 | -0.4941 | 0.0513 |
| 8 | `facebook/seamless-m4t-v2-large` | transformers_pipeline | 0.2079 | 0.0961 | 0.2848 | 0.9609 | 0.0527 | -0.3199 | 0.0391 |
| 9 | `openai/whisper-medium` | openai-whisper | 0.2143 | 0.1069 | 0.2151 | 0.9476 | 0.0791 | -0.3452 | 0.0498 |
| 10 | `mpoyraz/wav2vec2-xls-r-300m-cv8-turkish` | transformers_ctc | 0.2903 | 0.0665 | 0.1256 | 0.9418 | 0.0550 | -0.0986 | 0.0582 |
| 11 | `openai/whisper-small` | openai-whisper | 0.3000 | 0.1327 | 0.1178 | 0.9448 | 0.0638 | -0.2163 | 0.0543 |
| 12 | `facebook/mms-1b-fl102` | transformers_ctc | 0.4612 | 0.1454 | 0.3744 | 0.9297 | 0.0915 | -0.2346 | 0.0657 |
| 13 | `ceyda/wav2vec2-large-xlsr-53-turkish` | transformers_ctc | 0.4832 | 0.1118 | 0.1254 | 0.9320 | 0.0618 | -0.1680 | 0.0680 |
| 14 | `m3hrdadfi/wav2vec2-large-xlsr-turkish` | transformers_ctc | 0.4847 | 0.1102 | 0.1236 | 0.9316 | 0.0603 | -0.1159 | 0.0684 |
| 15 | `cahya/wav2vec2-large-xlsr-turkish` | transformers_ctc | 0.5218 | 0.1404 | 0.1554 | 0.9323 | 0.0604 | -0.1190 | 0.0677 |
| 16 | `mbsouksu/wav2vec2-large-xlsr-turkish-large` | transformers_ctc | 0.5224 | 0.1266 | 0.2073 | 0.9293 | 0.0648 | -0.1799 | 0.0707 |
| 17 | `openai/whisper-base` | openai-whisper | 0.5238 | 0.2291 | 0.1726 | 0.9324 | 0.0663 | -0.2435 | 0.0676 |
| 18 | `selimc/whisper-large-v3-turbo-turkish` | transformers | 0.5382 | 0.3481 | 0.0374 | 0.9510 | 0.0604 | -0.1126 | 0.0490 |
| 19 | `Baybars/wav2vec2-xls-r-300m-cv8-turkish` | transformers_ctc | 0.5550 | 0.1393 | 0.2130 | 0.9297 | 0.0634 | -0.1478 | 0.0703 |
| 20 | `erenfazlioglu/whisper-small-turkish-tr-best` | transformers | 0.6195 | 0.3500 | 0.0843 | 0.9515 | 0.0580 | -0.1228 | 0.0485 |
| 21 | `openai/whisper-tiny` | openai-whisper | 0.6547 | 0.3154 | 0.1083 | 0.9180 | 0.0859 | -0.1910 | 0.0802 |
| 22 | `gorkemgoknar/wav2vec2-large-xlsr-53-turkish` | transformers_ctc | 0.6759 | 0.2074 | 0.1230 | 0.9242 | 0.0633 | -0.1272 | 0.0758 |

Per-row clip counts are listed in the Clips column. The machine-readable table,
including runtime and source counts, is in `summary/leaderboard.csv`. Per-source
WER, CER, and RTF are in `summary/source_breakdown.csv`.

## AcoSemantic-TR

The release also includes ASCS_text, an experimental text-level semantic-affective
consistency score. It is reported as an additional analysis dimension and must not
be interpreted as a replacement for WER or CER.

For each clip:

```text
ASCS_text = 1 - |sentiment(reference) - sentiment(hypothesis)|
```

Sentiment scores were produced with
`lxyuan/distilbert-base-multilingual-cased-sentiments-student`. A value near `1`
means the reference and hypothesis received similar sentiment scores. This value
can remain high even when the hypothesis contains word errors, which is why ASCS
is shown alongside WER and CER.

`turkmedstt/whisper-large-v3-turkish-general` has the best WER and the highest
mean ASCS_text. Among the other public checkpoints, `facebook/seamless-m4t-v2-large`
has the highest mean ASCS_text (0.9609).

### SER domain-mismatch diagnostic

The earlier audio-level speech-emotion experiment is reported separately because
it did not produce a valid model-ranking metric on neutral read speech.

| Diagnostic | Result |
|---|---:|
| Evaluated clips | 199 |
| Mean maximum emotion confidence | 0.255 |
| Maximum confidence | 0.270 |
| Clips with confidence above 0.30 | 0 |
| Matched/mismatched AUC | 0.446 |

The acted-speech SER model collapsed to near-uniform predictions on Common Voice
and ISSAI read speech. This negative result motivated the text-based ASCS metric.

## Files

- `summary/leaderboard.csv`: overall WER, CER, RTF, runtime, source counts, and
  model-level AcoSemantic metrics.
- `summary/source_breakdown.csv`: metrics by model and source.
- `summary/acosemantic_summary.csv`: ASCS mean, standard deviation, WER
  correlation, sentiment means, and sentiment drift.
- `data/per_utterance_metrics.csv`: de-identified clip-level WER, CER, RTF, and timing.
- `data/acosemantic_per_utterance.csv`: de-identified clip-level WER, reference
  sentiment, hypothesis sentiment, and ASCS_text values.

Audio identifiers are retained only to join metrics across tables. No audio or
transcript content is included.

### Machine-readable field groups

| File | Important fields |
|---|---|
| `leaderboard.csv` | `rank`, `model`, `backend`, `clips`, `mean_wer`, `mean_cer`, `mean_rtf`, `mean_ascs_text`, `std_ascs_text`, `ascs_wer_correlation`, `mean_sentiment_drift` |
| `source_breakdown.csv` | `model`, `source`, `clips`, `mean_wer`, `mean_cer`, `mean_rtf` |
| `per_utterance_metrics.csv` | `model`, `audio_id`, `source`, `wer`, `cer`, `rtf`, `runtime_seconds`, `audio_seconds` |
| `acosemantic_per_utterance.csv` | `model`, `audio_id`, `source`, `wer`, `ref_sentiment`, `hyp_sentiment`, `ascs_text` |

## Scope and limitations

- These results cover general-domain Turkish speech only.
- Results are hardware- and implementation-dependent, especially RTF.
- Some original model attempts required dependency or loader repairs before the
  final valid runs.
- Source datasets remain governed by their original licenses and terms.
- This metrics release does not grant rights to redistribute the source audio or
  transcripts.
- Macro averaging gives short and long clips equal weight; corpus-level edit-count
  aggregation can produce different WER/CER values.
- Sentiment models can contain language and domain biases. ASCS should be treated as
  experimental supporting evidence.
- The leaderboard is a snapshot of tested checkpoints and software versions, not a
  permanent ranking of every Turkish ASR model.

## Contributors

- [Muhammed Kumcu](https://huggingface.co/muhammedkumcu)
- [Yagmur Tuncer](https://huggingface.co/yagmurtuncer)

Muhammed Kumcu and Yagmur Tuncer jointly carried out the benchmark design, dataset
preparation, model execution, metric analysis, documentation, and publication work.

## Acknowledgements

We thank Zeynep Zehra Kumcu and Yusuf Uysal for their support.

## Citation

```bibtex
@misc{kumcu_tuncer_2026_turkish_asr_benchmark,
  title        = {TurkMedSTT Turkish ASR Benchmark},
  author       = {Kumcu, Muhammed and Tuncer, Yagmur},
  year         = {2026},
  publisher    = {Hugging Face},
  organization = {TurkMedSTT}
}
```