didi-lot1-ai/ai_platform/modules/audio/API.md

# Audio Transcription API Documentation

OpenAI-compatible speech-to-text transcription API using faster-whisper.

## Base URL

```
{BASE_URL}
```

- **Local development:** `http://localhost:8200`
- **Docker (internal):** `http://audio-api:8200`
- **Direct:** `http://localhost:54300`
- **Production:** Use your configured hostname

## Authentication

No authentication required by default. Can be added via nginx or API gateway if needed.

## API Endpoints

### Health Check

Check API health status.

**Endpoint:** `GET /health`

**Response:**
```json
{
  "status": "ok"
}
```

**Example:**
```bash
curl http://localhost:8200/health
```

---

### List Models

List available Whisper models (OpenAI-compatible).

**Endpoint:** `GET /v1/models`

**Response:**
```json
{
  "object": "list",
  "data": [
    {
      "id": "large-v3-turbo",
      "object": "model",
      "created": 1700000000,
      "owned_by": "openai"
    }
  ]
}
```

**Example:**
```bash
curl http://localhost:8200/v1/models
```

---

### Create Transcription

Transcribe audio file to text (OpenAI-compatible endpoint).

**Endpoint:** `POST /v1/audio/transcriptions`

**Content-Type:** `multipart/form-data`

**Request Parameters:**

| Parameter | Type | Required | Default | Description |
|-----------|------|----------|---------|-------------|
| `file` | file | Yes | - | Audio file to transcribe (MP3, WAV, M4A, etc.) |
| `model` | string | No | `large-v3-turbo` | Model to use (currently ignored, uses configured model) |
| `language` | string | No | `null` | Language code (ISO-639-1). Auto-detected if not specified. |
| `prompt` | string | No | `null` | Optional text to guide the model's style |
| `response_format` | string | No | `json` | Format: `json`, `text`, or `verbose_json` |
| `temperature` | float | No | `0.0` | Sampling temperature (0.0-1.0). Use 0.0 for deterministic output. |

**Supported Languages (ISO-639-1 codes):**
`en`, `es`, `fr`, `de`, `it`, `pt`, `nl`, `pl`, `tr`, `ru`, `ja`, `ko`, `zh`, `ar`, `hi`, and 90+ more languages.

**Response Formats:**

#### 1. JSON (default)
```json
{
  "text": "Full transcription text",
  "language": "en",
  "duration": 45.5
}
```

#### 2. Text
```
Full transcription text
```

#### 3. Verbose JSON
```json
{
  "text": "Full transcription text",
  "language": "en",
  "duration": 45.5,
  "segments": [
    {
      "id": 0,
      "seek": 0,
      "start": 0.0,
      "end": 3.5,
      "text": "Hello, world!",
      "tokens": [15496, 11, 1002, 0],
      "temperature": 0.0,
      "avg_logprob": -0.25,
      "compression_ratio": 1.5,
      "no_speech_prob": 0.01
    }
  ]
}
```

**Examples:**

**Basic transcription (JSON):**
```bash
curl -X POST "http://localhost:8200/v1/audio/transcriptions" \
  -F "file=@audio.mp3" \
  -F "response_format=json"
```

**With language specification:**
```bash
curl -X POST "http://localhost:8200/v1/audio/transcriptions" \
  -F "file=@audio.mp3" \
  -F "language=en" \
  -F "response_format=json"
```

**Text format:**
```bash
curl -X POST "http://localhost:8200/v1/audio/transcriptions" \
  -F "file=@audio.mp3" \
  -F "response_format=text"
```

**Verbose JSON with segments:**
```bash
curl -X POST "http://localhost:8200/v1/audio/transcriptions" \
  -F "file=@audio.mp3" \
  -F "response_format=verbose_json"
```

**With initial prompt (to guide style):**
```bash
curl -X POST "http://localhost:8200/v1/audio/transcriptions" \
  -F "file=@audio.mp3" \
  -F "prompt=This is a technical discussion about machine learning." \
  -F "response_format=json"
```

**Python example:**
```python
import requests

url = "http://localhost:8200/v1/audio/transcriptions"

with open("audio.mp3", "rb") as f:
    files = {"file": f}
    data = {
        "model": "large-v3-turbo",
        "language": "en",
        "response_format": "verbose_json"
    }

    response = requests.post(url, files=files, data=data)
    result = response.json()

    print(f"Transcription: {result['text']}")
    print(f"Duration: {result['duration']}s")
    print(f"Segments: {len(result['segments'])}")
```

**JavaScript example:**
```javascript
const formData = new FormData();
formData.append('file', audioFile);
formData.append('model', 'large-v3-turbo');
formData.append('response_format', 'json');

const response = await fetch('http://localhost:8200/v1/audio/transcriptions', {
  method: 'POST',
  body: formData
});

const result = await response.json();
console.log('Transcription:', result.text);
```

---

## Error Responses

### Standard Error Format

```json
{
  "detail": "Error message"
}
```

### Common HTTP Status Codes

| Code | Meaning | Description |
|------|---------|-------------|
| 200 | OK | Request successful |
| 400 | Bad Request | Invalid request parameters |
| 413 | Payload Too Large | File exceeds max size limit (default 500MB) |
| 422 | Unprocessable Entity | Invalid file format or corrupted audio |
| 500 | Internal Server Error | Transcription failed |
| 503 | Service Unavailable | Model not loaded or GPU unavailable |

**Example error response:**
```json
{
  "detail": "File too large: 550.0MB (max: 500MB)"
}
```

---

## Rate Limiting

No built-in rate limiting. Can be added via nginx or API gateway.

**Recommended nginx configuration:**
```nginx
limit_req_zone $binary_remote_addr zone=audio_limit:10m rate=10r/m;

location / {
    limit_req zone=audio_limit burst=5;
    proxy_pass http://audio-api:8200;
}
```

---

## File Size Limits

- **Default:** 500MB per file
- **Configurable via:** `AUDIO_MAX_FILE_SIZE_MB` environment variable
- **Recommended:** Keep files under 100MB for best performance

---

## Supported Audio Formats

All formats supported by FFmpeg, including:

**Audio files:**
- MP3, WAV, FLAC, OGG, M4A, AAC, WMA, OPUS

**Video files (audio track extraction):**
- MP4, AVI, MKV, MOV, WEBM, FLV

**Professional formats:**
- PCM, AIFF, AU, AMR

---

## Performance Considerations

### Processing Time

Approximate transcription times for large-v3-turbo with int8 on H200 GPU:

| Audio Length | Processing Time | Real-Time Factor |
|--------------|-----------------|------------------|
| 1 minute | ~3-5 seconds | 0.05-0.08x |
| 10 minutes | ~30-50 seconds | 0.05-0.08x |
| 1 hour | ~3-5 minutes | 0.05-0.08x |

**Real-Time Factor (RTF):** Processing time / Audio duration. Lower is better.
- RTF < 0.1x = Excellent (8x faster than real-time)
- RTF < 0.5x = Good (2x faster than real-time)
- RTF = 1.0x = Real-time

### Optimization Tips

1. **Use language parameter** when known (skip auto-detection)
2. **Use appropriate model size** based on accuracy needs
3. **Process shorter segments** for long recordings (split at silence)
4. **Monitor GPU utilization** with `nvidia-smi`

---

## Comparison with OpenAI API

This API is **fully compatible** with OpenAI's transcription endpoint:

| Feature | OpenAI | This API |
|---------|--------|----------|
| Endpoint | `/v1/audio/transcriptions` | `/v1/audio/transcriptions` ✅ |
| Request format | `multipart/form-data` | `multipart/form-data` ✅ |
| Response formats | `json`, `text`, `verbose_json` | `json`, `text`, `verbose_json` ✅ |
| Language detection | Auto | Auto ✅ |
| Max file size | 25MB | 500MB (configurable) ✅ |
| Cost | $0.006/minute | Free (self-hosted) ✅ |
| Latency | Variable | ~0.05-0.08x RTF ✅ |

**Drop-in replacement:** Change only the base URL to switch from OpenAI to this API.

---

## SDK Examples

### OpenAI Python SDK

```python
from openai import OpenAI

# Point to local API
client = OpenAI(
    api_key="not-needed",  # No auth required
    base_url="http://localhost:8200/v1"
)

with open("audio.mp3", "rb") as f:
    transcript = client.audio.transcriptions.create(
        model="large-v3-turbo",
        file=f,
        response_format="verbose_json"
    )

print(transcript.text)
```

### OpenAI Node.js SDK

```javascript
import OpenAI from 'openai';
import fs from 'fs';

const openai = new OpenAI({
  apiKey: 'not-needed',
  baseURL: 'http://localhost:8200/v1'
});

const transcription = await openai.audio.transcriptions.create({
  file: fs.createReadStream('audio.mp3'),
  model: 'large-v3-turbo',
  response_format: 'verbose_json'
});

console.log(transcription.text);
```

---

## Advanced Features

### Voice Activity Detection (VAD)

Enabled by default. Automatically skips silence for:
- Faster processing
- Better accuracy
- Smaller output

### Beam Search

Configurable via `AUDIO_BEAM_SIZE` (default: 5). Higher values = better accuracy but slower.

### Temperature Sampling

- `0.0` (default): Deterministic output
- `0.0-1.0`: More creative/diverse outputs (less reliable)

---

## Monitoring

### Health Check

```bash
# Check if API is ready
curl http://localhost:8200/health

# Expected response
{"status": "ok"}
```

### Logs

```bash
# View real-time logs
cd deploy/
docker compose logs -f audio-api

# Check for errors
docker compose logs audio-api | grep ERROR
```

### GPU Usage

```bash
# Monitor GPU while transcribing
watch -n 1 nvidia-smi

# Check VRAM usage
nvidia-smi --query-gpu=memory.used,memory.total --format=csv
```

---

## Troubleshooting

### Error: File too large

**Solution:** Increase max file size or split audio into smaller chunks:
```bash
# Increase limit
AUDIO_MAX_FILE_SIZE_MB=1000

# Or split with ffmpeg
ffmpeg -i long_audio.mp3 -f segment -segment_time 600 -c copy chunk_%03d.mp3
```

### Error: Transcription failed

**Possible causes:**
1. Corrupted audio file - verify with media player
2. Unsupported format - convert to MP3/WAV
3. GPU out of memory - use smaller model or CPU mode
4. Audio is pure noise/music - Whisper is designed for speech

### Slow performance

**Check:**
1. GPU is being used: `docker exec audio-api nvidia-smi`
2. Model is correct: Check logs for model loading messages
3. Compute type is int8: Faster than float16

---

For more information, see [README.md](README.md) and the [faster-whisper documentation](https://github.com/SYSTRAN/faster-whisper).