keystats/kiswahili_sahihi_asr

🗣️ Kiswahili Sahihi ASR — Swahili Audio Transcription

This model enables high-quality, long-form Kiswahili speech transcription from multiple audio formats (e.g., .mp3, .wav, .m4a, .aac, .ogg, .flac, .amr) using a simple, efficient pipeline.
It’s optimized for speed, accuracy, and real-world usability, even on modest hardware.

---

🚀 Key Features

• ✅ Supports multiple audio formats via FFmpeg + Pydub
• 🧠 Built on 🤗 Transformers
• 🪶 Automatically converts audio to 16 kHz mono
• ⏳ Transcribes long recordings using smart chunking (default: 60s per chunk)
• 🖥️ Works seamlessly on both CPU and GPU
• 🌍 Focused on Kiswahili language transcription

🦊Example using the model

---

# ============================================
# 🪄 Full Swahili Audio Transcription Script 
# ============================================
# 📦 Install 
!pip install transformers
!pip install "datasets<4.0.0"
!pip install torchvision==0.21.0 torchaudio==2.6.0 jiwer evaluate
!pip install soundfile librosa accelerate>=0.26.0 tensorboard -U bitsandbytes
!apt-get -y install ffmpeg

import torch
import librosa
import numpy as np
from pydub import AudioSegment
from transformers import AutoProcessor, AutoModelForSpeechSeq2Seq
import os

# =============================
# 1. 🔸 Model Setup
# =============================
model_id = "keystats/kiswahili_sahihi_asr"
processor = AutoProcessor.from_pretrained(model_id)

# Use float32 to avoid half precision mismatch issues
model = AutoModelForSpeechSeq2Seq.from_pretrained(model_id).to("cuda", dtype=torch.float32)

# =============================
# 2. 🔸 Convert any format to WAV
# =============================
def convert_to_wav(input_path, output_path="converted.wav"):
    try:
        audio = AudioSegment.from_file(input_path)
        audio = audio.set_frame_rate(16000).set_channels(1)
        audio.export(output_path, format="wav")
        return output_path
    except Exception as e:
        raise RuntimeError(f"❌ Could not convert file. Check if FFmpeg is installed and file is supported. Error: {e}")

# 👇 Just change this path to your audio file
audio_path = "your swahili audio "
wav_path = convert_to_wav(audio_path)

# =============================
# 3. 🔸 Load audio and chunk
# =============================
audio_input, sr = librosa.load(wav_path, sr=16000, mono=True)

chunk_length_s = 60  # seconds
chunk_size = chunk_length_s * sr
num_chunks = int(np.ceil(len(audio_input) / chunk_size))

print(f"🔹 Total length: {len(audio_input)/sr:.2f} sec | Splitting into {num_chunks} chunks...")

# =============================
# 4. 🔸 Transcribe each chunk
# =============================
full_transcription = []

for i in range(num_chunks):
    start = i * chunk_size
    end = min((i + 1) * chunk_size, len(audio_input))
    chunk = audio_input[start:end]

    inputs = processor(
        chunk,
        sampling_rate=16000,
        return_tensors="pt",
        padding=True
    ).to("cuda", dtype=torch.float32)

    with torch.no_grad():
        generated_ids = model.generate(**inputs, max_length=20000)

    text = processor.batch_decode(generated_ids, skip_special_tokens=True)[0]
    full_transcription.append(text.strip())

# =============================
# 5. 🔸 Combine final transcript
# =============================
final_text = " ".join(full_transcription)
print(" 📝 Final Transcription:")
print(final_text)

---

🧪 Example Output

🎧 Input Audio	📝 Transcription Output
mashairi_sauti.mp3	“Karibu kwenye mfumo wetu wa Kiswahili Sahihi.”
mazungumzo_flac.flac	“Habari yako, karibu tena kesho kwa mahojiano mengine.”

---

🛠️ Tips for Best Results

• Use clear audio without background noise.
• Long recordings are automatically split into 60-second chunks.
• Works with .mp3, .wav, .m4a, .aac, .ogg, .flac, .amr and more.
• Ensure audio is sampled at 16 kHz and mono (automatically handled).

---

🌟 Acknowledgements

• 🤗 Transformers
• Mozilla Foundation
• Zindi Africa

---

📢 Contribute

• 🧪 Share more Swahili audio samples
• 🧑‍💻 Report issues or improvements
• 🌍 Help expand coverage for different accents and dialects

---

🧭 Citation

---

@model{kiswahili_sahihi_asr,
  author    = {Jackson Kahungu},
  title     = {Kiswahili Sahihi ASR — Swahili Audio Transcription},
  year      = {2025},
  publisher = {Hugging Face}
}

---

---

✨ Final Note

“If you like the model, leave a like ❤🧡❤”
This model may not be perfect, but it provides a strong baseline for building future Swahili transcription systems.
Together, we can make Swahili voice technology accessible to everyone.✨ 🎊KISWAHILI KITUKUZWE🎉