hasan0v

whisper-turbo-az

README

License: apache-2.0

Accurate Azerbaijani transcription that runs on your own GPU.

Whisper Turbo AZ is a full fine-tune of Whisper large-v3-turbo for Azerbaijani speech-to-text. It continues from LocalDoc/azerbaijani-whisper-turbo, which was already strong, and lowers its word error rate on FLEURS-az test from 13.11% → 10.90%. That is a 16.9% relative reduction.

It writes cased, punctuated text. It was trained to use the previous sentence as context, which helps on long recordings such as meetings, lectures and podcasts. It also has a CTranslate2 / faster-whisper export that transcribes about 28× faster than real time on a laptop RTX 4070.

Table

Use it for	Interviews, meetings, lectures, podcasts, voice notes, subtitles, voice interfaces
Input	16 kHz mono audio, any length (long audio is chunked in 30 s windows)
Output	Azerbaijani text in Latin script, cased and punctuated
Formats	🤗 Transformers (this repo) · CTranslate2 fp16 for faster-whisper (`whisper-turbo-az-ct2`) · ONNX for transformers.js / WebGPU (`whisper-turbo-az-onnx`)
Try it	Live browser demo: runs on your own device, no upload

Results

FLEURS-az test, all 921 utterances. Every model was scored with the same script: faster-whisper, fp16, beam 5, no VAD, language="az". Text is normalised the same way as in the LocalDoc evaluation: NFC, lowercase, punctuation removed.

Table with columns: Model, WER ↓, CER ↓, Relative WER
Model	WER ↓	CER ↓	Relative WER
LocalDoc/azerbaijani-whisper-turbo (starting point)	13.11%	3.39%	—
Whisper Turbo AZ v1 (filtered data + context training)	11.72%	3.06%	−10.6%
Whisper Turbo AZ v2 (this model)	10.90%	2.86%	−16.9%

Head-to-head with a commercial API

This comparison uses the first 200 FLEURS-az test clips with the same normalisation. OpenAI gpt-transcribe was called through its API.

Table with columns: Model, WER ↓, CER ↓, Where it runs
Model	WER ↓	CER ↓	Where it runs
OpenAI gpt-transcribe	10.70%	3.14%	Paid cloud API
Whisper Turbo AZ v2	11.64%	2.85%	Your GPU, offline, free
Whisper Turbo AZ v1	12.35%	2.99%	Your GPU, offline, free
LocalDoc turbo	13.42%	3.24%	Your GPU, offline, free

On character error rate, v2 makes 9% fewer errors than gpt-transcribe. Its word error rate is within one point of the API. The low CER suggests that many of the remaining word errors are small spelling or suffix differences, not missing words.

Sample transcriptions (FLEURS-az test)

Table with columns: Text
	Text
Reference	sundarban 1966-cı ildən bəri vəhşi təbiət qoruğudur ərazidə indi 400 benqal pələnginin …
LocalDoc turbo	sundurban 1966-cı ildən bəri vəhşi təbiət qorxmuşdur ərazidə indi 400 benqal pələnginin …
v2	sundarban 1966-cı ildən bəri vəhşi təbiət qoruğundur ərazidə indi 400 benqal pələnginin …
Reference	… ev pişiklərinin böcəklər gəmiricilər kərtənkələlər və quşlar kimi …
LocalDoc turbo	… ev pişiklərinin böcəklər gəmircilər kərtənkələrlər və quşlar kimi …
v2	… ev pişiklərinin böcəklər gəmiricilər kartankələrlər və quşlar kimi …

Quick start

🤗 Transformers

python

import torchfrom transformers import pipeline asr = pipeline(    "automatic-speech-recognition",    model="hasan0v/whisper-turbo-az",    torch_dtype=torch.float16,    device="cuda:0",          # or "cpu") result = asr(    "audio.wav",    generate_kwargs={"language": "az", "task": "transcribe"},    chunk_length_s=30,        # long-form audio    batch_size=8,)print(result["text"])

⚡ faster-whisper (recommended for production)

python

from faster_whisper import WhisperModel model = WhisperModel("hasan0v/whisper-turbo-az-ct2", device="cuda", compute_type="float16") segments, info = model.transcribe(    "audio.wav",    language="az",    beam_size=5,    vad_filter=True,                    # skip silence    condition_on_previous_text=True,    # the model was trained to use this context)for s in segments:    print(f"[{s.start:6.1f} → {s.end:6.1f}] {s.text}")

On CPU, use device="cpu", compute_type="int8". For high throughput on long files, use faster_whisper.BatchedInferencePipeline.

How it was built

Label cleaning. The starting model transcribed all 345,643 clips (329.6 h) of LocalDoc/azerbaijani_asr. Clips whose reference disagreed with that transcript by more than 30% CER were dropped: 11,733 clips, or 5.4 h. Most of them were sub-2-second fragments with broken alignments.
Context packing. Consecutive segments from the same source recording were merged into windows of up to 29.5 s. In 30% of samples, the previous window's text was given as a <|startofprev|> prompt. The result was 70,407 windows totalling 338.4 h, with a median length of 17 s. About 20% of windows were cut to shorter random lengths so the model still handles short utterances well.
Training v1. A full fine-tune on one H100 for 2,200 steps.
Extra data. v1 was used as a teacher:
- Audiobooks (asmarhajizada/azerbaijani-audiobooks): only segments where v1 agrees with the book text (CER ≤ 0.2) were kept, giving 25.1 h. The previous text comes from the same book chapter.
- YODAS2 az000: v1 pseudo-labelled the audio with VAD and beam 5. Segments passed only if avg_logprob > −0.5, no_speech_prob < 0.5 and compression_ratio < 2.2, and any output in Cyrillic script was rejected. This gave 10.9 h.
- FLEURS-az train: 9.25 h.
Training v2. Continued from v1's best checkpoint for 900 steps, with the new data repeated twice.
Leak check. Every training window was compared against all 491 normalised FLEURS test and dev sentences. 0 matches were found in any source.

The early rise in dev WER during v1 is the usual effect of learning-rate warm-up on a model that is already strong. v1 passes the starting point at around step 1,300 and keeps improving. v2 picks up from v1's best checkpoint and reaches 11.99% dev WER.

Hyperparameters

Table with columns: v1, v2
	v1	v2
Initialisation	LocalDoc/azerbaijani-whisper-turbo	v1 best checkpoint
Data	LocalDoc ASR (filtered, packed)	v1 data + audiobooks, YODAS2, FLEURS train (×2)
Steps	2,200	900
Effective batch	64 (16 × 4 grad. accumulation)	64
Peak LR / schedule	1e-5, cosine, 220 warm-up	5e-6, cosine, 90 warm-up
Optimiser	AdamW (β = 0.9 / 0.98, wd 0.01), grad-clip 1.0	same
Precision	bf16 autocast	bf16 autocast
Augmentation	SpecAugment (time 0.05, freq 0.05)	same
Context prompt	previous text, p = 0.3, ≤ 200 tokens	same
Checkpoint selection	best FLEURS-az dev WER, evaluated every 250 steps	same
Hardware	1 × NVIDIA H100 80 GB, about 1.6 s/step	same

The whole project cost under $10 of cloud GPU. That covers data filtering, packing, pseudo-labelling and both training runs.

Limitations

The model is tuned for Azerbaijani in Latin script. Other languages, and Azerbaijani written in Cyrillic or Arabic script (South Azerbaijani), are out of scope.
Heavy dialects, strong code-switching with Russian, Turkish or English, and noisy far-field audio will have higher error rates than FLEURS read speech.
Numbers may come out as digits or as words, depending on context.
Like every Whisper model, it can occasionally hallucinate text during long silences or music. vad_filter=True in faster-whisper avoids most of this.
Do not use it as the only source for high-stakes decisions, such as legal or medical transcripts, without human review.

Training data and licences

Table with columns: Source, Hours used, Notes
Source	Hours used	Notes
LocalDoc/azerbaijani_asr	338.4	CER-filtered, context-packed
asmarhajizada/azerbaijani-audiobooks	25.1	Filtered by agreement with v1
espnet/yodas2 (`az000`)	10.9	Pseudo-labelled, CC-BY 3.0
LocalDoc/fleurs-azerbaijani-asr (train)	9.25	CC-BY 4.0

The model weights are released under Apache 2.0, the same licence as the base model. Please check each dataset's own terms if you use this model commercially.

Azərbaycanca

Whisper Turbo AZ Azərbaycan dilində nitqi mətnə çevirən modeldir. O, Whisper large-v3-turbo əsasında qurulub və LocalDoc modelindən davam etdirilib. FLEURS-az test dəstində WER 13.11%-dən 10.90%-ə endirilib. Model mətni böyük-kiçik hərflə və durğu işarələri ilə yazır. Uzun audioda əvvəlki cümləni kontekst kimi istifadə edir. Öz GPU-nuzda, internetsiz və pulsuz işləyir. Sürətli inference üçün whisper-turbo-az-ct2 versiyasını faster-whisper ilə istifadə edin.

Citation

bibtex

@misc{hasanov2026whisperturboaz,  title        = {Whisper Turbo AZ: Azerbaijani Speech Recognition with Whisper large-v3-turbo},  author       = {Hasanov, Ali},  year         = {2026},  howpublished = {\url{https://huggingface.co/hasan0v/whisper-turbo-az}}}

Thanks to LocalDoc for the base model and datasets, to the FLEURS, YODAS2 and Azerbaijani audiobooks contributors, and to OpenAI for Whisper.

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Model Details

Model Provider

hasan0v

Model Tree

Base

LocalDoc/azerbaijani-whisper-turbo

Fine-tuned

this model

Input Modalities

Audio

Output Modalities

Text

Supported Functionality

Dedicated Endpoints

Explore FriendliAI today

Get started Talk to an engineer

README

License: apache-2.0

Accurate Azerbaijani transcription that runs on your own GPU.

Table

Use it for	Interviews, meetings, lectures, podcasts, voice notes, subtitles, voice interfaces
Input	16 kHz mono audio, any length (long audio is chunked in 30 s windows)
Output	Azerbaijani text in Latin script, cased and punctuated
Formats	🤗 Transformers (this repo) · CTranslate2 fp16 for faster-whisper (`whisper-turbo-az-ct2`) · ONNX for transformers.js / WebGPU (`whisper-turbo-az-onnx`)
Try it	Live browser demo: runs on your own device, no upload

Results

Table with columns: Model, WER ↓, CER ↓, Relative WER
Model	WER ↓	CER ↓	Relative WER
LocalDoc/azerbaijani-whisper-turbo (starting point)	13.11%	3.39%	—
Whisper Turbo AZ v1 (filtered data + context training)	11.72%	3.06%	−10.6%
Whisper Turbo AZ v2 (this model)	10.90%	2.86%	−16.9%

Head-to-head with a commercial API

This comparison uses the first 200 FLEURS-az test clips with the same normalisation. OpenAI gpt-transcribe was called through its API.

Table with columns: Model, WER ↓, CER ↓, Where it runs
Model	WER ↓	CER ↓	Where it runs
OpenAI gpt-transcribe	10.70%	3.14%	Paid cloud API
Whisper Turbo AZ v2	11.64%	2.85%	Your GPU, offline, free
Whisper Turbo AZ v1	12.35%	2.99%	Your GPU, offline, free
LocalDoc turbo	13.42%	3.24%	Your GPU, offline, free

On character error rate, v2 makes 9% fewer errors than gpt-transcribe. Its word error rate is within one point of the API. The low CER suggests that many of the remaining word errors are small spelling or suffix differences, not missing words.

Sample transcriptions (FLEURS-az test)

Table with columns: Text
	Text
Reference	sundarban 1966-cı ildən bəri vəhşi təbiət qoruğudur ərazidə indi 400 benqal pələnginin …
LocalDoc turbo	sundurban 1966-cı ildən bəri vəhşi təbiət qorxmuşdur ərazidə indi 400 benqal pələnginin …
v2	sundarban 1966-cı ildən bəri vəhşi təbiət qoruğundur ərazidə indi 400 benqal pələnginin …
Reference	… ev pişiklərinin böcəklər gəmiricilər kərtənkələlər və quşlar kimi …
LocalDoc turbo	… ev pişiklərinin böcəklər gəmircilər kərtənkələrlər və quşlar kimi …
v2	… ev pişiklərinin böcəklər gəmiricilər kartankələrlər və quşlar kimi …

Quick start

🤗 Transformers

python

import torchfrom transformers import pipeline asr = pipeline(    "automatic-speech-recognition",    model="hasan0v/whisper-turbo-az",    torch_dtype=torch.float16,    device="cuda:0",          # or "cpu") result = asr(    "audio.wav",    generate_kwargs={"language": "az", "task": "transcribe"},    chunk_length_s=30,        # long-form audio    batch_size=8,)print(result["text"])

import torchfrom transformers import pipeline asr = pipeline(    "automatic-speech-recognition",    model="hasan0v/whisper-turbo-az",    torch_dtype=torch.float16,    device="cuda:0",          # or "cpu") result = asr(    "audio.wav",    generate_kwargs={"language": "az", "task": "transcribe"},    chunk_length_s=30,        # long-form audio    batch_size=8,)print(result["text"])

⚡ faster-whisper (recommended for production)

python

from faster_whisper import WhisperModel model = WhisperModel("hasan0v/whisper-turbo-az-ct2", device="cuda", compute_type="float16") segments, info = model.transcribe(    "audio.wav",    language="az",    beam_size=5,    vad_filter=True,                    # skip silence    condition_on_previous_text=True,    # the model was trained to use this context)for s in segments:    print(f"[{s.start:6.1f} → {s.end:6.1f}] {s.text}")

from faster_whisper import WhisperModel model = WhisperModel("hasan0v/whisper-turbo-az-ct2", device="cuda", compute_type="float16") segments, info = model.transcribe(    "audio.wav",    language="az",    beam_size=5,    vad_filter=True,                    # skip silence    condition_on_previous_text=True,    # the model was trained to use this context)for s in segments:    print(f"[{s.start:6.1f} → {s.end:6.1f}] {s.text}")

On CPU, use device="cpu", compute_type="int8". For high throughput on long files, use faster_whisper.BatchedInferencePipeline.

How it was built

Label cleaning. The starting model transcribed all 345,643 clips (329.6 h) of LocalDoc/azerbaijani_asr. Clips whose reference disagreed with that transcript by more than 30% CER were dropped: 11,733 clips, or 5.4 h. Most of them were sub-2-second fragments with broken alignments.
Context packing. Consecutive segments from the same source recording were merged into windows of up to 29.5 s. In 30% of samples, the previous window's text was given as a <|startofprev|> prompt. The result was 70,407 windows totalling 338.4 h, with a median length of 17 s. About 20% of windows were cut to shorter random lengths so the model still handles short utterances well.
Training v1. A full fine-tune on one H100 for 2,200 steps.
Extra data. v1 was used as a teacher:
- Audiobooks (asmarhajizada/azerbaijani-audiobooks): only segments where v1 agrees with the book text (CER ≤ 0.2) were kept, giving 25.1 h. The previous text comes from the same book chapter.
- YODAS2 az000: v1 pseudo-labelled the audio with VAD and beam 5. Segments passed only if avg_logprob > −0.5, no_speech_prob < 0.5 and compression_ratio < 2.2, and any output in Cyrillic script was rejected. This gave 10.9 h.
- FLEURS-az train: 9.25 h.
Training v2. Continued from v1's best checkpoint for 900 steps, with the new data repeated twice.
Leak check. Every training window was compared against all 491 normalised FLEURS test and dev sentences. 0 matches were found in any source.

Hyperparameters

Table with columns: v1, v2
	v1	v2
Initialisation	LocalDoc/azerbaijani-whisper-turbo	v1 best checkpoint
Data	LocalDoc ASR (filtered, packed)	v1 data + audiobooks, YODAS2, FLEURS train (×2)
Steps	2,200	900
Effective batch	64 (16 × 4 grad. accumulation)	64
Peak LR / schedule	1e-5, cosine, 220 warm-up	5e-6, cosine, 90 warm-up
Optimiser	AdamW (β = 0.9 / 0.98, wd 0.01), grad-clip 1.0	same
Precision	bf16 autocast	bf16 autocast
Augmentation	SpecAugment (time 0.05, freq 0.05)	same
Context prompt	previous text, p = 0.3, ≤ 200 tokens	same
Checkpoint selection	best FLEURS-az dev WER, evaluated every 250 steps	same
Hardware	1 × NVIDIA H100 80 GB, about 1.6 s/step	same

The whole project cost under $10 of cloud GPU. That covers data filtering, packing, pseudo-labelling and both training runs.

Limitations

The model is tuned for Azerbaijani in Latin script. Other languages, and Azerbaijani written in Cyrillic or Arabic script (South Azerbaijani), are out of scope.
Heavy dialects, strong code-switching with Russian, Turkish or English, and noisy far-field audio will have higher error rates than FLEURS read speech.
Numbers may come out as digits or as words, depending on context.
Like every Whisper model, it can occasionally hallucinate text during long silences or music. vad_filter=True in faster-whisper avoids most of this.
Do not use it as the only source for high-stakes decisions, such as legal or medical transcripts, without human review.

Training data and licences

Table with columns: Source, Hours used, Notes
Source	Hours used	Notes
LocalDoc/azerbaijani_asr	338.4	CER-filtered, context-packed
asmarhajizada/azerbaijani-audiobooks	25.1	Filtered by agreement with v1
espnet/yodas2 (`az000`)	10.9	Pseudo-labelled, CC-BY 3.0
LocalDoc/fleurs-azerbaijani-asr (train)	9.25	CC-BY 4.0

The model weights are released under Apache 2.0, the same licence as the base model. Please check each dataset's own terms if you use this model commercially.

Azərbaycanca

Citation

bibtex

@misc{hasanov2026whisperturboaz,  title        = {Whisper Turbo AZ: Azerbaijani Speech Recognition with Whisper large-v3-turbo},  author       = {Hasanov, Ali},  year         = {2026},  howpublished = {\url{https://huggingface.co/hasan0v/whisper-turbo-az}}}

@misc{hasanov2026whisperturboaz,  title        = {Whisper Turbo AZ: Azerbaijani Speech Recognition with Whisper large-v3-turbo},  author       = {Hasanov, Ali},  year         = {2026},  howpublished = {\url{https://huggingface.co/hasan0v/whisper-turbo-az}}}

Thanks to LocalDoc for the base model and datasets, to the FLEURS, YODAS2 and Azerbaijani audiobooks contributors, and to OpenAI for Whisper.