whisper-small-tt-ggml

Whisper Small · Tatar · GGML q4_0

A 4-bit whisper.cpp build of yasalma/whisper-finetuned-tt-asr, a Tatar fine-tune of openai/whisper-small. It runs fully offline on a laptop CPU/GPU: 139 MB, about a third of a second per short clip on an Apple M4.

Made for Tatarby, a Tatar-language voice assistant.

Usage

hf download JackGauter/whisper-small-tt-ggml ggml-tt-small-q4_0.bin --local-dir models

# one file
whisper-cli -m models/ggml-tt-small-q4_0.bin -l tt -f speech.wav

# HTTP server: POST /inference (multipart field `file`) → {"text": "..."}
whisper-server -m models/ggml-tt-small-q4_0.bin --language tt \
  --host 127.0.0.1 --port 8178 --threads 4 \
  --suppress-nst --no-timestamps --flash-attn --convert

Set the language explicitly (-l tt / --language tt) — the fine-tune was trained for Tatar only, and the evaluation below used this setting. --convert lets the server accept MP3/OGG/etc. through ffmpeg; without it, send 16 kHz mono WAV.

Evaluation

First 100 clips of yasalma/TatSC_ASR, config crowdsourced_web, split test. Text is NFC-normalized, lower-cased, ё→е, punctuation removed before scoring. whisper.cpp server with the flags above, 4 threads, Apple M4.

4-bit quantization costs about 2 WER points against f16 while being 3.3× smaller. 100 clips is a small sample, so treat differences under ~2 points as noise.

How it was made

1. Downloaded model.safetensors and the tokenizer/config files from yasalma/whisper-finetuned-tt-asr (safetensors SHA-256 68254ad4272a85d977f37f161c7580c2378df8ff42210dd795e176c962c552c0).
2. Generated vocab.json and added_tokens.json from tokenizer.json, because convert-h5-to-ggml.py expects the legacy slow-tokenizer files.
3. Converted to GGML f16 with models/convert-h5-to-ggml.py from whisper.cpp commit 1da4dc82fa7996d4edda05890dca65aeceaafd6d, using the mel filters from openai/whisper commit 86098128c0b4f24f0e2aa2994de830614b474227.
4. Quantized with whisper-quantize ggml-tt-small-f16.bin ggml-tt-small-q4_0.bin q4_0 from the same whisper.cpp commit.

Limitations

• Inherits the base model's training data (TatSC_ASR, tat_hackathon_asr, Common Voice — mostly read speech); expect lower accuracy on noisy audio, spontaneous speech and Tatar–Russian code-switching.
• Like other Whisper models it can hallucinate text on silence or music; trim silence or use VAD for long recordings.
• For best accuracy use the original model or a higher-precision conversion.

License and credits

Apache-2.0, same as the base model.

• Fine-tuned model and datasets: yasalma
• Original architecture and weights: OpenAI Whisper (MIT)
• Conversion and inference: whisper.cpp (MIT)