Instructions to use derkar00/whisper-base-turkmen with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use derkar00/whisper-base-turkmen with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("automatic-speech-recognition", model="derkar00/whisper-base-turkmen")# pip install -U transformers accelerate # Load model directly from transformers import AutoProcessor, AutoModelForSpeechSeq2Seq processor = AutoProcessor.from_pretrained("derkar00/whisper-base-turkmen") model = AutoModelForSpeechSeq2Seq.from_pretrained("derkar00/whisper-base-turkmen", device_map="auto") - Notebooks
- Google Colab
- Kaggle
Whisper Base β Turkmen v2 (conversational + Common Voice)
Full fine-tune of OpenAI Whisper Base for Turkmen (74 M parameters), used by the ses. voice-translation app. FP32 SafeTensors with the Transformers configuration. Fast CTranslate2 INT8 version: derkar00/whisper-base-turkmen-ct2.
This is v2. v1 (checkpoint 8,500, trained only on TurkmenSpeech) stays available at the git tag
checkpoint-8500; use--revision checkpoint-8500. v2 is also taggedv2-conversational. Pin a tag or commit for reproducible results.
Results (v2, independent test sets)
These benchmarks are independent of this model's training data. None of the recordings or speakers below was used to train v2.
| Model / decoding | Conversational test, 170 clips (WER / CER %) | Common Voice 27.0 tk test, 521 clips (WER / CER %) |
|---|---|---|
v2, beam 5 + repetition guard (default in infer.py) |
62.24 / 27.63 | 42.17 / 10.59 |
| v2, greedy + repetition guard | 63.13 / 28.86 | 45.84 / 11.78 |
| v1 checkpoint 8,500, greedy | 71.3 / 30.7 | 52.9 / 14.3 |
| Reference: Pirnazar/whisper-large-v3-turbo-turkmen-ct2 (809 M), beam 5 | 54.61 / 22.31 | 40.22 / 8.65 |
- Conversational test: 23 minutes of casual, multi-speaker Turkmen from two YouTube recordings that were never used for training. References were produced with Gemini assistance and human-verified. This set is private and not redistributed.
- Common Voice 27.0 Turkmen, official test split: 3 sentences that also occur in TurkmenSpeech were removed. v2 was trained on the separate CV train split; its 2 speakers do not appear in the test split.
- Clean broadcast speech (TurkmenSpeech, 7,470 held-out clips, greedy): v1 25.9 / 5.9 β v2 25.4 / 5.9 WER / CER. These clips are held out from both versions' Whisper training.
- Scoring: text is lowercased, NFC-normalized and reduced to Turkmen letters before scoring. Measured with the production CTranslate2 int8_float16 runtime on an RTX 3050.
- Median latency per clip: about 0.11β0.16 s (beam 5) and 0.11β0.14 s (greedy).
- Beam 5 makes the repetition guard trigger slightly more often on fast conversational speech (23 vs 17 of 170 clips).
- Expect high error rates on spontaneous conversation. About 60% WER is still a hard-audio result. A larger model (see the Pirnazar row) remains about 8 WER points better on conversation.
Installation and inference
Use Python 3.12. For CPU-only PyTorch, install its CPU wheel first:
python -m venv .venv
source .venv/bin/activate
python -m pip install "torch==2.11.0" --index-url https://download.pytorch.org/whl/cpu
python -m pip install "transformers==4.57.1" "numpy==1.26.4" "soundfile==0.13.1" "scipy==1.15.3"
hf download derkar00/whisper-base-turkmen infer.py --local-dir turkmen-whisper
python turkmen-whisper/infer.py speech.wav --device cpu # beam 5 + repetition guard
python turkmen-whisper/infer.py speech.wav --device cpu --beam 1 # greedy
WAV/FLAC input is mixed to mono and resampled to 16 kHz. Longer files are processed in 30-second windows, including the tail. --json adds timing, beam and repetition-guard information; --revision pins a tag or commit.
from infer import Transcriber # run from the downloaded folder
asr = Transcriber("derkar00/whisper-base-turkmen", device="cpu", threads=4, beam=5)
print(asr.transcribe("speech.wav")["text"])
What changed from v1 (checkpoint 8,500)
v2 starts from v1 (OpenAI Whisper Base fully fine-tuned on mamed0v/TurkmenSpeech) and was fine-tuned further in two stages. All parameters were trained in BF16 with an effective batch of 32.
- Conversational stage: 1,350 updates on 2,413 clips (3.66 h) of conversational Turkmen from 14 sources (YouTube recordings and private recordings).
- Transcripts were produced with Google Gemini from audio plus an ASR draft. About a third were human-verified by the publisher; the rest were automatically quality-checked.
- Each clip got speed (0.9/1.1), gain and noise augmentation.
- Each epoch also replayed 3Γ as many random TurkmenSpeech training clips, to limit forgetting.
- Common Voice stage: 1,120 updates on the same conversational data plus Common Voice 27.0 Turkmen train (719 clips, 1.0 h, CC0, repeated twice per epoch), again with TurkmenSpeech replay. LR 5e-6.
- Checkpoint selection: only on development sets (99 conversational dev clips, the 516-clip CV dev split, 256 TurkmenSpeech selection clips). Clean-speech WER was not allowed to worsen by more than 1 point.
- Unchanged: the tokenizer, the decoding prompt and the architecture.
- Not redistributed: no training audio or transcripts are included in this repository.
Decoding: keep the two-token prompt
Like v1, this model was trained with the prompt <|startoftranscript|><|notimestamps|> ([50258, 50363]), not the standard four-token <|tk|><|transcribe|> prompt. Forcing the language and task tokens makes the model loop badly. infer.py keeps the correct prompt and adds the repetition guard used in production: a window whose gzip compression ratio exceeds 2.4 is re-decoded once with repeated 3-grams blocked.
Limitations and license
- Conversational accuracy is still modest. Fast, overlapping, noisy or dialect-heavy speech produces many errors; review important transcripts.
- Output is lowercase Turkmen Latin without punctuation (the training text style).
- No speaker or demographic balance analysis has been done. The conversational data comes from a small number of sources.
- License: CC BY-NC 4.0 (non-commercial), inherited from TurkmenSpeech. See
LICENSEandNOTICE.md. Upstream OpenAI Whisper notices are inlicenses/.
- Downloads last month
- 89